<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Yash Gorasiya on Medium]]></title>
        <description><![CDATA[Stories by Yash Gorasiya on Medium]]></description>
        <link>https://medium.com/@yashgorasiya?source=rss-835e0d12af1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Uj7aPGukzo7TgCt-QMWkkQ.jpeg</url>
            <title>Stories by Yash Gorasiya on Medium</title>
            <link>https://medium.com/@yashgorasiya?source=rss-835e0d12af1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 12:00:05 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yashgorasiya/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI Is Making Decisions. Your Controls Aren’t.]]></title>
            <link>https://yashgorasiya.medium.com/ai-is-making-decisions-your-controls-arent-56222db351ea?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/56222db351ea</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 30 Mar 2026 08:31:01 GMT</pubDate>
            <atom:updated>2026-03-30T08:31:01.052Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nIzIQkd1_DM3epW78qvdQA.png" /></figure><p><em>Welcome back. Pour yourself something warm. This one takes a minute to sit with.</em></p><p>There is a version of this conversation happening in boardrooms right now that goes roughly like this:</p><p>“Are our AI systems secure?”</p><p>“Yes. Access controls are in place. Data is encrypted. Vendors are assessed.”</p><p>“Good.”</p><p>And then the meeting moves on.</p><p>That exchange is not wrong, exactly. But it is answering the wrong question. What nobody is asking, at least not consistently, is whether those AI systems are making good decisions. Whether those decisions are drifting. Whether anyone would know if they weren’t.</p><p>We have spent years building out our security posture around a relatively stable idea of what risk looks like. Who has access to what. Whether systems are up. Whether data has moved somewhere it shouldn’t. That framing made sense when the technology we were protecting was passive. Databases don’t act on their own. Servers don’t choose. Applications do what they’re told.</p><p>AI doesn’t quite work that way.</p><p>It infers. It weights. It prioritises. It acts. And increasingly, it does all of this inside systems that your security team has certified as compliant, monitored, and under control.</p><p>The system is secured. The decisions it makes may not be.</p><p>That is the gap Edition 26 is about.</p><h3>The Layer We’re Protecting vs The Layer That’s Actually Moving</h3><p>Let’s be precise about this, because the conflation is causing real confusion.</p><p>Traditional security controls are built around a well-understood stack. You protect the infrastructure. You manage identity and access. You monitor the network. You classify and secure data. You harden endpoints. The threat model here is largely adversarial: someone is trying to get in, take something, or disrupt something. The controls you build are designed to prevent or detect that kind of intervention.</p><p>This is not a criticism. That stack is necessary. It works. It keeps a lot of very bad things from happening.</p><p>But AI systems introduce a layer that sits above all of that. The decision layer. And that layer doesn’t behave like anything the traditional security model was built to protect.</p><p>Take a fraud detection model. Infrastructure: secured. Data pipelines: encrypted. Access: role-based and audited. The model itself: approved, tested, deployed. From a security posture perspective, everything looks fine.</p><p>Now ask: is the model making good fraud decisions today? Is it flagging the same proportion of genuine transactions as suspicious that it was six months ago? Has the distribution of cases it’s trained on drifted away from the cases it’s now seeing in production? Is it systematically more aggressive toward a particular transaction type because of something subtle in the data it was trained on?</p><p>None of those questions are answered by your SIEM. None of them show up in your access logs. None of them trip a firewall rule.</p><p>They show up, eventually, as customer complaints. Regulatory questions. Operational friction. Business loss. By then, the model has been making those decisions for weeks or months.</p><p>The same pattern plays out across almost every domain where AI is being deployed at scale. Credit scoring systems that quietly became less accurate after the economy shifted. Hiring filters that screen out candidates in ways that would not survive a legal challenge if anyone looked closely. SOC alert prioritisation tools that are routing the wrong signals to the top of the queue because the threat landscape changed and the model didn’t. Customer service automation that is resolving tickets in ways that satisfy the metric but not the customer.</p><p>These are not exotic scenarios. They are happening now, in organisations with mature security programmes, because the security programme was never designed to catch them.</p><h3>What Happens When Decisions Become the Risk</h3><p>There is something worth sitting with here, because it challenges a deeply ingrained assumption about what risk looks like.</p><p>In security, we are conditioned to look for events. Something happens: a login attempt, a file transfer, an anomalous network connection. Something triggers. Someone investigates. That event-driven model of risk detection has shaped everything from our tooling to our team structures to the mental models we bring to risk conversations.</p><p>AI decision risk doesn’t move like that.</p><p>It doesn’t arrive as an event. It accumulates. Quietly, across thousands or millions of decisions, a model’s behaviour gradually diverges from what was originally intended. The shift might be almost imperceptible in any single decision. But at scale, and over time, the aggregate effect is significant.</p><p><strong>Model drift</strong> is the formal term for this. It happens when the statistical properties of the data a model encounters in production start to diverge from the data it was trained on. The world changes. Consumer behaviour shifts. Fraud patterns evolve. Threat actors adapt. The model, frozen in its training, doesn’t.</p><p>Some drift is slow. A gradual erosion of accuracy. The model gets slightly less reliable each month. If nobody is looking, nobody knows. If the business metric the model is contributing to is also drifting, it may take a long time to even notice the correlation.</p><p>Some drift is faster, triggered by a sharp external change. A new regulation alters what transactions look like. A major event reshapes how people behave online. A supply chain disruption changes purchasing patterns. The model, trained on pre-event data, is suddenly operating in an environment it wasn’t built for. But it keeps running. Keeps making decisions. Nobody turned it off.</p><p>Then there are feedback loops. This one is underappreciated. When a model’s decisions influence the next round of data it’s trained on, errors can reinforce themselves. A credit model that underweights certain applicants leads to fewer approvals in that group, which leads to less repayment data for that group, which leads to the model becoming even less calibrated for them over time. The model doesn’t know it’s doing this. Neither does anyone watching the dashboard that shows it’s performing within acceptable parameters.</p><p>None of this looks like a security incident. That is precisely the problem.</p><h3>Why the Controls You Have Don’t Help Here</h3><p>Let’s go through this directly, because there’s a tendency to assume that good security hygiene covers more than it actually does.</p><p>Access control tells you who can touch a system, not what the system is doing with the access it has. A model running with appropriate service account permissions is, from an IAM perspective, behaving correctly. It doesn’t matter if the decisions it’s generating are wrong.</p><p>Logging gives you a record of what happened. It does not give you a record of why. You can log every API call, every model invocation, every input and output. But what you capture is transactional data. You don’t capture reasoning. You don’t capture the features the model weighted most heavily. You don’t capture what the model would have decided if one input had been slightly different. Without that, your audit trail tells you a decision was made, but it cannot tell you whether that decision was sound.</p><p>SIEM platforms are built to detect patterns that signal threat activity. Anomalous login times. Unusual data volumes. Known attack signatures. They are not built to detect “the model’s false negative rate on a particular transaction type has increased by 12% over the last 90 days.” That is not a threat signal. It’s a model performance signal. Different domain entirely.</p><p>Security testing, even thorough penetration testing and red team exercises, evaluates whether a system can be exploited. It does not evaluate whether the system’s core decision-making function is working correctly under normal operating conditions. A model that passes a security review may still be biased, brittle, or drifting. The review was never designed to catch that.</p><p>What we have, if we are honest about it, is deep visibility into who accessed what and whether systems are running as expected. We have almost no visibility into what decisions were made and whether those decisions were right.</p><p>That gap is not a tooling problem you can solve by buying something. It’s an architectural problem. It’s a governance problem. The controls were designed around a different threat surface.</p><h3>The Visibility Gap Is Worse Than Most People Think</h3><p>Here is where I want to slow down, because this is the part that doesn’t get enough attention.</p><p>Decision-level telemetry, meaning the systematic monitoring of what decisions an AI system is making, how those decisions are distributed, how they are changing over time, is genuinely hard. Not insurmountably hard, but hard enough that most organisations haven’t built it.</p><p>Part of the difficulty is definitional. For a firewall, “normal behaviour” is relatively straightforward to characterise. For a machine learning model making complex inferences, defining what “normal” looks like requires understanding the model’s intended behaviour, the distribution of inputs it was designed for, the acceptable range of outcomes, and how that range might legitimately shift over time. That requires collaboration between security, data science, and the business owners of whatever the model is supporting. It requires someone who understands both risk management and model behaviour. That person exists in very few organisations today.</p><p>Part of the difficulty is volume. A large model deployed in production might be making millions of decisions a day. You cannot review them manually. You need automated systems that can detect distributional shifts in outputs, flag unexpected clusters of decisions, and alert on behaviour that falls outside established bounds. Those systems need to exist, need to be calibrated, and need to be monitored by someone who can actually interpret what they’re showing.</p><p>Part of the difficulty is the black box problem. For some AI systems, particularly deep learning models, the relationship between inputs and outputs is not transparent even to the people who built them. Explainability techniques like SHAP or LIME can offer partial insight, but they are approximations. They tell you something about what the model is likely doing, not a definitive account of its reasoning. That matters a great deal when you are trying to audit whether a consequential decision was made for the right reasons.</p><p>Detection without understanding is not enough. You can detect that a model’s outputs have shifted. You cannot always explain why, or predict what conditions will cause it to shift further. That limits your ability to manage the risk intelligently.</p><h3>The Third-Party Problem: Delegating Decisions You Can’t See</h3><p>Everything above applies to models your organisation built and operates directly. Now layer in third-party AI.</p><p>Most organisations are not running their own foundation models. They are consuming AI capabilities through APIs, through SaaS platforms with embedded AI features, through vendor products that use machine learning under the hood. The enterprise software you rely on for HR, finance, security operations, customer management: a significant proportion of it now has AI making or influencing decisions inside it.</p><p>What visibility do you have into how those decisions are made?</p><p>In most cases: very little. You have a contract. You have a SLA. You might have a data processing agreement. You have a vendor security assessment that checks whether their infrastructure is appropriately protected. You do not have visibility into the model’s training data, its drift monitoring practices, its explainability posture, or the criteria it uses to make consequential recommendations.</p><p>This is supply chain risk at the decision layer. The framing most people use for supply chain risk is about code integrity and software components. That framing is important but incomplete. When you integrate an AI vendor, you are not just integrating their software. You are integrating their decision-making. Their model’s assumptions and biases and limitations become, in practice, your organisation’s assumptions and biases and limitations. And you are accountable for the outcomes even when the underlying logic is opaque to you.</p><p>This is not theoretical. An organisation using an AI-powered recruitment tool is, in a real legal sense, responsible for the fairness of the hiring decisions that tool influences. An organisation using an AI credit decisioning platform carries accountability for those credit outcomes even if the model itself is the vendor’s intellectual property. Outsourcing the decision doesn’t outsource the liability.</p><p>The security team’s current approach to vendor AI is, in most organisations, an extension of the standard vendor assessment. Is the data secure? Is access controlled? Is the infrastructure compliant? Those are necessary questions. They are not sufficient ones.</p><h3>What Regulators Are Starting to Expect</h3><p>The regulatory landscape is moving in a direction that makes this gap progressively harder to ignore.</p><p>The EU AI Act, which came into force in 2024 with obligations phasing in through 2026 and 2027, establishes a risk-based classification framework. High-risk AI systems, those used in areas like employment, credit, critical infrastructure management, and law enforcement, carry specific requirements around transparency, human oversight, accuracy, and ongoing monitoring. The accountability expectation is explicit: organisations deploying these systems must be able to demonstrate that the systems work as intended, that risks are being managed, and that meaningful human oversight exists.</p><p>In the UK, the approach has been more principles-based and sector-led, with regulators like the FCA, ICO, and the Bank of England developing their own AI governance guidance within their domains. The direction is consistent with the EU in one important respect: accountability sits with the organisation deploying the AI, not the organisation that built it. The question regulators are starting to ask is not just “Is this system secure?” It is “Can you demonstrate this system is making appropriate decisions, and how do you know?”</p><p><strong>That is a very different question.</strong> And most organisations, if they are honest, cannot fully answer it today.</p><p>The gap between what regulators are beginning to expect and what most governance frameworks currently deliver is real. The organisations that will find themselves in difficulty are not necessarily the ones with the weakest security posture. They are the ones that continued to conflate system security with decision accountability.</p><h3>My View on Where We Are Getting This Wrong</h3><p>Let me be direct about something that comes up often in conversations with security and governance leaders.</p><p>There is a comfortable assumption operating underneath a lot of AI governance discussions, which is that model accuracy is a reasonable proxy for decision quality. If the model performs well on its benchmark metrics, if accuracy is above some threshold, if the business unit that deployed it is satisfied with the results, then the governance box is ticked.</p><p>That assumption is wrong. Not slightly wrong. Dangerously wrong.</p><p>Accuracy metrics are measured against historical test sets. They tell you how the model performed on data it was evaluated on, at the time it was evaluated. They tell you almost nothing about how it is performing today, on the actual distribution of cases it is encountering, in the context of how the world has changed since the test set was assembled. A model can carry a 95% accuracy score from its last formal evaluation while currently operating at 78% on a materially different real-world distribution. Nobody has recalculated. Nobody will know until something goes wrong.</p><p>Beyond accuracy, there is the question of what “correct” even means for a consequential decision. A fraud model might correctly classify a transaction as low risk based on all available signals and still be wrong, because the signals themselves were incomplete or the adversary adapted. A hiring filter might apply its criteria consistently and still encode historical patterns that no reasonable person would endorse if stated explicitly. Accuracy at the model level and appropriateness at the outcome level are not the same thing.</p><p>I also think organisations are underestimating how quickly the AI decision surface is expanding. Three years ago, AI was influencing a handful of specific functions in most enterprises. Today it is embedded in security tooling, in HR platforms, in financial controls, in customer operations, in supply chain management. Each of those is a domain where AI is now making or heavily influencing decisions that have real consequences for real people. The governance frameworks being applied to those systems were largely built before AI was operating at this scale and this depth. They haven’t caught up.</p><p>And boards. I say this with genuine respect for the difficulty of board-level oversight, but most board AI governance discussions I am aware of are still anchored to two concerns: is the data secure, and is there reputational risk from a high-profile failure. Those are legitimate concerns. They are not the full picture. The more important questions, how do we know the AI decisions embedded in our operations are sound, what would drift look like before it became a visible problem, who in this organisation owns decision-level accountability, are not yet being asked consistently. That needs to change.</p><p>The organisations that are doing this well are the ones where someone with genuine AI and risk fluency has been empowered to connect the dots between data science, security, governance, and business outcomes. They are a minority. The gap between what they understand and what the average enterprise AI governance programme covers is significant.</p><h3>What Actually Needs to Change</h3><p>This is not a comprehensive framework. It is a direction. The specifics will vary by organisation, sector, and maturity. But there are a few things that need to shift fundamentally.</p><p><strong>CISOs and security leaders need to extend their accountability horizon.</strong></p><p>This does not mean owning data science. It means accepting that “AI security” is not just about securing the systems AI runs on. It extends to the integrity of what those systems decide. That requires building relationships with the AI and data engineering teams that are actually running these systems, understanding what monitoring exists at the model layer, and identifying where decision risk sits outside anyone’s current accountability.</p><p>The security function is well-positioned to push for decision-level observability because it already understands what robust monitoring looks like. Apply that understanding to the model layer. What does a baseline look like? What constitutes an anomaly? Who gets alerted and when? Those are questions security teams are good at answering for infrastructure. Start applying them to decisions.</p><p><strong>Organisations need to define what acceptable decision behaviour looks like before something goes wrong.</strong></p><p>This is the equivalent of threat modelling, applied to AI systems. For each AI system making consequential decisions: what does good look like, what does bad look like, what would drift look like, and what is the mechanism for detecting it? Most organisations have not done this exercise. It requires convening people who don’t usually sit in the same room: data scientists, security, legal, compliance, and the business owners of the system. It is uncomfortable and time-consuming and genuinely necessary.</p><p>Decision thresholds need to exist. Not just performance thresholds from the model evaluation, but operational thresholds tied to business and risk outcomes. When those thresholds are breached, there needs to be a defined response. Right now, for most AI systems in production, there is no equivalent of the incident response plan. There is no defined trigger for “the model is behaving unexpectedly.” That is an accountability gap.</p><p><strong>Third-party AI needs to be treated with the same seriousness as any other outsourced critical function.</strong></p><p>That means going beyond infrastructure security assessment. Vendor due diligence for AI systems should include questions about drift monitoring practices, explainability posture, model update and change management processes, and contractual obligations around performance accountability over time. Some vendors will not be able to answer these questions well. That is useful information. It means you are operating in a governance gap, and you need to either close it or accept and document the residual risk consciously.</p><p><strong>Boards need to ask a different question.</strong></p><p>Not “are our AI systems secure?” But: “How do we know the decisions our AI systems are making are sound, and what would it look like if they weren’t?”</p><p>That question, asked persistently, will drive the right conversations. It will surface where accountability is unclear. It will expose where monitoring is inadequate. It will create pressure on the people who need to answer it to actually build the infrastructure to answer it.</p><h3>A Scenario Worth Thinking Through</h3><p>Consider this.</p><p>A mid-size financial services organisation deploys an AI system to support operational triage: routing customer queries, escalating cases, flagging potential compliance issues for human review. The system was evaluated rigorously before deployment. It performs well. The team is satisfied.</p><p>Twelve months pass. The macro environment shifts. Customer behaviour changes. The types of queries coming in evolve. The distribution of what “needs escalation” looks like in practice starts to diverge quietly from what it looked like in the training data.</p><p>The model doesn’t know this. It continues applying the patterns it learned. It escalates what it was trained to escalate. But some of what now genuinely needs human attention isn’t getting flagged, because it doesn’t look like the cases that were labelled “escalate” in 2022.</p><p>There are no security alerts. The system is running normally. Logs show expected throughput. The vendor dashboard shows the model operating within its stated performance parameters, which were set based on the original test set.</p><p>Six months later, a cluster of customer complaints surfaces. A regulatory query follows. An internal investigation finds that a category of cases had been systematically under-escalated for most of the preceding year.</p><p>Nobody made a bad decision, in the traditional sense. The model was doing what it was built to do. The infrastructure was secure. The access was controlled. But the decisions were wrong. Quietly, gradually, at scale.</p><p>And there was no control that would have caught it.</p><p><em>This scenario is an analytical illustration, not a documented real-world case.</em></p><h3>Where This Leaves Us</h3><p>We secured the perimeter. Then we secured the endpoint. Then we secured identity. Each time the threat surface shifted, the security discipline followed, eventually.</p><p>The threat surface has shifted again. It shifted the moment organisations started deploying AI systems that make consequential decisions autonomously or semi-autonomously at scale. That shift is not coming. It happened. The question is whether the governance and security disciplines follow fast enough to matter.</p><p>Decision integrity is not yet a mature discipline. There is no ISO standard specifically for it. There is no tool you can buy that solves it cleanly. The regulatory frameworks are directional but not yet prescriptive enough to force the right conversations at the right level. That ambiguity creates space for organisations to assume they are covered when they are not.</p><p>The uncomfortable reality is this: you probably have AI systems operating in your organisation right now that are making consequential decisions and where nobody has a clear answer to the question “how do we know those decisions are still good?” That’s not a failure of security. It’s a failure of scope. The scope of what we treat as a governance and risk responsibility hasn’t kept up with what AI is actually doing inside our organisations.</p><p>Securing the system is necessary. It is not sufficient.</p><p>The next frontier is not a technical one. It’s a governance one. It’s the question of whether we are willing to hold ourselves accountable not just for protecting what we’ve built, but for interrogating what it decides.</p><p>That accountability doesn’t have a clean home yet. Somewhere between the CISO’s office, the data science team, the risk function, and the board, it is currently nobody’s job in a way that is clear and enforceable.</p><p>Making it someone’s job, with real authority, real visibility, and real accountability, is the work.</p><p>Everything else is infrastructure.</p><p><em>The views in this article are my own and are intended to prompt reflection, not serve as legal, compliance, or regulatory advice. Always apply these ideas in the context of your organisation’s specific circumstances.</em></p><p><strong><em>Get connected with me here :</em></strong></p><p><strong><em>LinkedIn: </em></strong><a href="https://www.linkedin.com/in/yash-gorasiya/"><em>https://www.linkedin.com/in/yash-gorasiya/</em></a></p><p><strong><em>Twitter: </em></strong><a href="https://twitter.com/r3v3Ncl4W"><em>https://twitter.com/r3v3Ncl4W</em></a></p><p><strong><em>Instagram: </em></strong><a href="https://www.instagram.com/infosec.yash/"><em>https://www.instagram.com/infosec.yash/</em></a></p><p><strong><em>Blog page: </em></strong><a href="https://instagram.com/forensis_digitum"><em>https://instagram.com/forensis_digitum</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=56222db351ea" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI Incident Response Plan Doesn’t Exist. It Just Looks Like It Does.]]></title>
            <link>https://yashgorasiya.medium.com/your-ai-incident-response-plan-doesnt-exist-it-just-looks-like-it-does-48afaa833c06?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/48afaa833c06</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[incident-response]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 23 Mar 2026 09:31:00 GMT</pubDate>
            <atom:updated>2026-03-23T09:31:00.570Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Rp56bT9km5jXyGomVeg2tg.png" /></figure><p>You’ve probably got an incident response plan.</p><p>Maybe it’s been reviewed this year. Maybe it’s been tested. You have playbooks, escalation paths, severity tiers, a war room protocol. Someone owns the comms. Someone owns the forensics. The CISO knows who to call at 2am.</p><p>Now ask yourself, and be honest about it: where does an AI failure fit into that?</p><p>Not a breach involving an AI vendor. Not data exfiltration from an AI product. I mean an AI system your organisation is running, embedded in a workflow, making or influencing decisions, and something goes wrong. The outputs become unreliable. The model starts behaving differently than it did six months ago. A recommendation engine quietly shifts, and nobody noticed until a business unit raised it in a quarterly review.</p><p>Where does that go? Who declares it an incident? Who investigates? What does remediation even look like?</p><p>If you had to pause on any of those questions, that pause is the problem.</p><h3>The Comfort of Documentation</h3><p>There is a specific kind of organisational risk that lives behind completed documents. The risk of something looking ready when it isn’t. And in the world of AI governance, we are manufacturing this faster than almost any other domain I can think of.</p><p>Boards are approving AI policies. Legal teams are reviewing AI usage clauses. CISOs are tagging AI-related risks in GRC platforms. Compliance teams are mapping activities to frameworks like NIST AI RMF or preparing for EU AI Act obligations. On paper, things look increasingly serious.</p><p>Except the operational layer, the part that actually has to respond when something breaks, is running on infrastructure that was never designed for this.</p><p>Traditional incident response was built for a specific class of problem. A threat actor gets in. A vulnerability is exploited. Data moves where it shouldn’t. Systems go down. These scenarios share some fundamental characteristics: they are discrete, often detectable through technical telemetry, and they have a clear before and after. Something changed. Something happened. You can point to a log, a timestamp, an artefact.</p><p>AI failures don’t behave like that.</p><h3>How Traditional IR Thinks About Incidents</h3><p>It’s worth being precise here, because the structural mismatch is important.</p><p>When a traditional incident occurs, ransomware, for instance, the sequence is roughly: detection, triage, containment, eradication, recovery, lessons learned. You know something is wrong because systems behave in unexpected and measurable ways. Encryption events fire. Anomalous network traffic appears. EDR platforms generate alerts. Even in more subtle scenarios, like a business email compromise or insider threat, there are usually artefacts. Login anomalies. Unusual data access. Lateral movement patterns.</p><p>The entire detection and response apparatus in modern security operations, the SIEMs, the EDR platforms, the UEBA tools, the threat intelligence feeds, is built to surface these signals. They monitor logs, endpoints, network traffic, user behaviour. They are calibrated to the idea that an incident produces observable technical evidence.</p><p>Now consider what model drift looks like.</p><p>A machine learning model in production begins degrading. Maybe it was fine for eighteen months. Its training data is no longer representative of current inputs. The distribution of data it’s seeing has shifted. Its predictions are subtly less accurate, or its outputs have begun reflecting a bias that wasn’t visible at lower volumes. The model hasn’t crashed. No alert has fired. No log entry says “model performance degraded.” The system, from an infrastructure standpoint, is healthy. It’s processing requests and returning outputs exactly as it always has.</p><p>The problem is that what it’s returning has changed in ways that matter.</p><p>And here’s the uncomfortable part: in many organisations, that change would surface not as a security alert, but as a slightly off quarterly metric. Or a cluster of customer complaints. Or a product manager noticing something feels wrong in the outputs. Or a business decision that, in retrospect, was based on faulty model output for three months before anyone connected the dots.</p><p>That is an incident. But nobody called it one.</p><h3>The Gap Between Detection and Interpretation</h3><p>This is where the real structural problem sits, and it doesn’t get discussed enough.</p><p>Detection in cybersecurity, at its core, is about recognising that something unexpected has occurred. The tooling is designed to generate signals based on deviation from a known baseline. When a user authenticates from an unusual location, or an application starts communicating with an external IP it’s never contacted, a signal fires. The SOC investigates.</p><p>AI systems require a different kind of detection. You’re not asking “<strong>did something happen?</strong>” You’re asking “<strong>is what’s happening still correct?</strong>” Those are fundamentally different questions. The first is answered by monitoring infrastructure. The second requires understanding the system’s purpose, its expected behaviour, its acceptable output range, and continuously evaluating whether those things hold.</p><p>Most SOC teams aren’t equipped to do that. And this isn’t a criticism; it’s a structural reality. Security operations professionals are trained to investigate incidents, not to evaluate whether a credit risk model is making sound predictions or whether a document classification system has developed a systematic error. Those are data science and machine learning engineering questions. They live in a completely different part of the organisation, often operating with no integration into security monitoring whatsoever.</p><p>What you end up with is a detection gap that’s invisible on both sides. The security team doesn’t see the signal because the tooling doesn’t generate it. The data science team doesn’t flag it because they may not have real-time behavioural monitoring in place, or because the degradation is slow enough that it doesn’t trigger whatever automated checks they’ve built. The business teams feel it eventually, but they describe it in business language, not security language. And the translation never happens.</p><p>This is what I mean when I say the gap is between detection and interpretation. Even in organisations that do have some monitoring of AI systems, the ability to translate a behavioural anomaly into a meaningful incident classification, to say “this is a problem, this is the severity, this is who owns it”, barely exists.</p><p>MITRE ATLAS, the adversarial threat landscape framework for AI systems, has done useful work cataloguing attack vectors specific to machine learning. Techniques like model inversion, data poisoning, adversarial inputs. But what ATLAS describes is largely the adversarial end of the spectrum. The non-adversarial failures, drift, distributional shift, silent degradation, are not attack techniques. They’re operational failures with security and governance implications. And they fall into a gap that neither the traditional security frameworks nor the AI operations communities have fully bridged.</p><h3>Nobody Owns This</h3><p>When something goes wrong in a traditional incident, there is usually a reasonably clear ownership structure. The security team investigates. Legal gets notified. Communications prepares a statement if needed. HR is involved if the incident involves an insider. The CISO coordinates. There is a hierarchy of decision-making that has been rehearsed, documented, and tested.</p><p>Ask who owns an AI incident.</p><p>The security team owns security incidents, but is model drift a security incident? The data science team owns model performance, but are they part of your incident response process? The product team owns user-facing experience, but do they have a formal incident escalation pathway? The risk team owns operational risk, but do they have visibility into AI system behaviour at a level of granularity that would allow them to declare a threshold breach?</p><p>In most organisations, the answer to all of those is some variant of “sort of” or “it depends.” Which means when something actually goes wrong, the first thing that happens is not containment or investigation. It’s a conversation about who should be in the conversation.</p><p>That delay is not trivial. In a traditional incident, the first hours matter enormously. Contamination spreads. Evidence degrades. The window for effective containment closes. AI incidents have analogous dynamics. If a model has been producing problematic outputs for weeks, every downstream decision made on the basis of those outputs is potentially compromised. The longer the ownership question goes unresolved, the broader the impact.</p><p>Traditional incident response works in part because ownership is assumed. The security team is the first responder. Everything else follows from that. AI incident response has no equivalent default. The ambiguity is baked in. And organisations have not, in most cases, resolved it with any deliberate structural decision. They’ve just left it unresolved and called it good.</p><h3>The Third-Party Problem Nobody Talks About Loudly Enough</h3><p>So far, I’ve been talking largely about first-party AI failures, models that organisations have built or deployed internally. The third-party picture is considerably more concerning.</p><p>Most organisations are not building foundation models. They’re consuming AI through APIs. They’re integrating large language models into workflows through vendor platforms. They’re using AI-embedded SaaS products where the AI component is one feature among many. And in most of those cases, they have minimal visibility into what’s happening at the model level.</p><p>Consider a realistic scenario. An organisation integrates an LLM-based API into its customer service workflow. The vendor, a well-known foundation model provider, updates the model. This is normal, providers regularly release updated model versions, adjust safety filters, change output formatting, or modify system-level behaviour. The organisation’s API contract may not even guarantee notification of these changes.</p><p>After the update, the model’s outputs change in subtle ways. Not broken, just different. The tone shifts. Certain topics are handled differently. Edge cases that used to resolve one way now resolve another. If the organisation had a tight testing regime around that integration, they might catch it. Most don’t. The change propagates into production outputs, and nobody declares an incident because nobody knows one has occurred.</p><p>The AI Incident Database, maintained by Partnership on AI, has documented hundreds of cases where AI system failures caused real-world harm. What’s striking about many of them is not the failure itself, it’s the gap between when the failure began and when it was recognised as a failure. In some cases, systems exhibiting problematic behaviour continued operating for extended periods before anyone formally stopped them.</p><p>Third-party model dependencies create a supply chain risk that most IR frameworks have not adequately incorporated. Software supply chain security is a well-developed discipline, we know how to think about dependencies, vulnerability disclosure, version management. AI model dependencies require an equivalent discipline that simply hasn’t matured yet. You can’t inspect a model update the way you can inspect a software patch. You can’t diff the weights of a neural network and immediately understand what changed behaviourally. The opacity is structural.</p><p>The question organisations are not asking, and need to be asking, is what triggers an incident review when a vendor changes a model you depend on. In most cases the answer is nothing, unless the change causes an obvious, immediate failure. Subtle behavioural shifts go uninvestigated because there’s no mechanism to detect them and no process that requires anyone to look.</p><h3>Regulatory Expectations Are Outpacing Operational Reality</h3><p>The regulatory environment around AI is moving, unevenly, but directionally, toward accountability.</p><p>The EU AI Act, which came into force in 2024, establishes requirements for high-risk AI systems that include obligations around transparency, human oversight, logging, and post-market monitoring. For organisations operating high-risk systems, there will be expectations around detecting and reporting serious incidents. The Act’s definition of a serious incident includes reasonably foreseeable risks to health, safety, or fundamental rights.</p><p>The UK’s approach has been more principle-based through the existing sector regulators, the FCA, ICO, and CMA, each extending their existing frameworks to cover AI within their respective domains. The ICO has been clear that automated decision-making carries accountability obligations under UK GDPR. The FCA has issued guidance on model risk management that applies increasingly to AI-driven decisions in financial services. None of these frameworks require you to have a formal AI incident response process. But all of them create accountability expectations that assume one exists.</p><p>And this is where the gap becomes legally and reputationally significant. When a regulator asks, and increasingly they will ask, “what was your response to this AI failure?” the expected answer is not “we didn’t classify it as an incident.” The expected answer is a documented process of detection, investigation, impact assessment, and remediation. If you can’t provide that, you’re not just operationally unprepared. You have a governance failure on your hands.</p><p>NIST’s AI Risk Management Framework, released in January 2023, provides a voluntary but widely referenced structure for thinking about AI risk across the model lifecycle. Its GOVERN, MAP, MEASURE, and MANAGE functions create a vocabulary for AI risk that is increasingly being adopted in enterprise risk programmes. But having adopted that vocabulary, having mapped your AI systems in a GRC tool and labelled them against the AI RMF categories, doesn’t mean you’ve built the operational capability to respond when something breaks. The framework is a scaffold. The actual structure is something organisations have to build themselves. Most haven’t.</p><h3>Why Current Incident Response Fails for AI, Structurally</h3><p>Let me try to be concrete about the structural gaps, because they are individually navigable but collectively they form a system failure.</p><p>There is no AI-specific incident taxonomy in most organisations. Traditional IR has severity classifications, P1 through P4, or equivalent, tied to business impact criteria like systems affected, data exposed, services disrupted. Nobody has defined what constitutes a P1 AI incident. Is it a model producing harmful outputs at scale? Is it a biased output affecting a protected class? Is it undetected drift in a fraud detection model? Without a taxonomy, you can’t triage. Without triage, you can’t escalate. And without escalation, you don’t have incident response, you have ad hoc problem management.</p><p>There is no behavioural monitoring baseline for most deployed AI systems. In traditional security, establishing a baseline is day one. You need to know what normal looks like before you can identify abnormal. Most AI systems in production do not have equivalent baselines. There may be performance metrics, accuracy, recall, precision, but these are often sampled, not continuously monitored, and they’re not integrated into any alerting or incident framework. Nobody is watching for drift the way a SOC watches for network anomalies.</p><p>There are no response playbooks for AI failure modes. Think about how specific good IR playbooks are. A ransomware playbook tells you which systems to isolate, which forensic artefacts to preserve, which stakeholders to notify, how to assess data exfiltration, how to communicate with leadership. An AI incident playbook, if it existed, would need to tell you how to assess output impact across a time period, how to roll back or suspend a model safely, how to communicate uncertainty to stakeholders when the exact scope of harm is unclear, how to evaluate third-party model changes. Nobody has written that document. And the absence of it means the first person to respond to a real AI incident is improvising.</p><h3>My POV</h3><p>I’ll be direct.</p><p>The organisations I see taking AI governance seriously are, almost without exception, doing it from the top down. Board-level AI policies. CISO-owned AI risk registers. Participation in industry frameworks. And a lot of it is real, the intent is genuine. But the gap between governance-level seriousness and operational-level preparedness is vast, and I don’t think most leaders have fully reckoned with it.</p><p>The version of AI incident response that exists in most organisations right now is: “we’ll handle it when it happens.” And that sounds reasonable until you ask what “handling it” actually looks like on day one of a real AI failure. Who makes the first call? To whom? On what basis do they declare an incident? What do they preserve? What do they shut down? How do they assess impact? Who talks to the regulator?</p><p>“We’ll handle it” only works if the people who need to handle it have the training, the tools, and the authority to act. In most cases, they don’t. What they have is a general competence in handling the wrong class of problem.</p><p>There’s also a compliance theatre dimension emerging that genuinely worries me. Completing a framework mapping exercise is not the same as building incident response capability. Tagging an AI system as “high risk” in a GRC tool does not create the monitoring that would detect a failure in that system. Publishing an AI policy does not train a SOC analyst to recognise behavioural anomalies in a model. These things can coexist, the documentation and the operational gap, and they will coexist right up until the moment they need to not coexist.</p><p>The other thing I keep coming back to is this: AI incidents will first appear as business problems. A drop in model accuracy shows up as a change in customer satisfaction scores. A biased recommendation engine shows up as a demographic pattern in complaints. A drift in a risk-scoring model shows up as anomalies in downstream portfolio performance. These signals go to business analysts, product managers, and customer service leads, not to the security team. The path from “something seems off with this AI system” to “this is a declared incident requiring a structured response” is currently a path that most organisations haven’t paved.</p><p>By the time something lands on the CISO’s desk as an AI incident, it’s usually already a business incident with AI at the root. That inversion matters. It means the response window is shorter than it appears, and the impact is already broader than initially visible.</p><h3>What Actually Needs to Change</h3><p>This isn’t a checklist. But there are some concrete shifts that would meaningfully reduce the gap.</p><p><strong><em>For CISOs</em></strong>: The most urgent thing is definitional. Your incident classification framework needs AI-specific categories before you have an incident that requires them. Define what a model drift event looks like, what an unsafe output event looks like, what a third-party AI change event looks like, and map those to your existing severity tiers. This doesn’t need to be perfect, it needs to exist, because even an imperfect taxonomy gives your team something to anchor to. From there, expand your IR plan to include explicit sections on AI system failures, with named owners from data science and product, not just security.</p><p><strong><em>For organisations:</em></strong> Behavioural monitoring for production AI systems isn’t optional anymore, it’s a foundational control. This means building or procuring the capability to track not just whether a model is running, but what it’s producing. Output quality metrics, distribution monitoring, human-in-the-loop review sampling for high-stakes systems. And those monitoring outputs need to route into someone who has the authority and the process to escalate them. The integration between AI operations and security operations is not a nice-to-have. It’s the gap where incidents live.</p><p>Shared responsibility models for third-party AI also need to be formalised. When you integrate an AI API, you need contractual commitments around change notification, incident reporting, and model versioning. You need testing gates that run on production-equivalent inputs before AI vendor updates go live. And you need someone who owns the question “did the model change, and if so, how?”</p><p><strong><em>For boards:</em></strong> The questions you should be asking in AI risk reviews are not “do we have an AI policy?” and “are we compliant with relevant frameworks?” Those questions have been answered, or will be soon. The harder questions are: do we have the operational capability to detect an AI failure before a business metric flags it? Do we know who declares an AI incident, and under what criteria? Have we tested our AI incident response in the last twelve months? If the answers are unclear, that’s the work.</p><h3>A Scenario Worth Walking Through</h3><p>Imagine a mid-sized financial services firm. It has deployed a machine learning model to assist relationship managers with client risk profiling, not to make decisions autonomously, but to generate recommendations that managers review and approve. The model has been in production for two years. It performs well on its original training distribution.</p><p>Over the course of about six months, the nature of the firm’s client base shifts. New client segments are onboarded. The input data the model receives starts diverging from what it was trained on, not dramatically, but consistently. The model’s recommendations start reflecting patterns that were relevant eighteen months ago, not today. The recommendations are still being approved by relationship managers, most of whom trust the system and move through approvals quickly.</p><p>Three months into this drift, a compliance analyst notices an unusual pattern in portfolio allocations across a client segment. It doesn’t look wrong, exactly, just slightly inconsistent with what she’d expect. She raises it informally. It gets noted. A few weeks later, the pattern is more visible. An internal audit team picks it up. By the time it’s formally investigated, the model has been in a degraded state for around five months.</p><p>The response is improvised. Nobody has a playbook. The data science team is brought in but isn’t part of the security or risk response structure. Legal and compliance have questions that nobody knows how to answer. The firm doesn’t know how to scope the impact, which client recommendations were affected, to what degree, over what period. The communication to regulators is delayed while this is being worked out internally. The reputational damage is contained, but the operational response is messy, expensive, and leaves everyone aware that they were not prepared for this class of problem.</p><p><em>This scenario is an analytical illustration, not a documented real-world case.</em></p><h3>Closing</h3><p>The gap I’ve been describing isn’t theoretical. It’s being quietly accumulated right now, in organisations with mature security programmes, experienced CISOs, and genuinely good intent.</p><p>The problem is not that people don’t take AI risk seriously. It’s that the seriousness is being expressed in governance language, in framework language, in documentation language, while the operational layer, the part that actually has to respond at 11pm when something breaks, remains largely unprepared for the specific class of failure that AI systems produce.</p><p>We are preparing for the wrong kind of incidents.</p><p>And when the right kind arrives, quiet, slow, ambiguous, spreading through business metrics before it ever becomes a security alert, the playbook we reach for will be the wrong one.</p><p><strong><em>Get connected with me here :</em></strong></p><p><strong><em>LinkedIn: </em></strong><a href="https://www.linkedin.com/in/yash-gorasiya/"><em>https://www.linkedin.com/in/yash-gorasiya/</em></a></p><p><strong><em>Twitter: </em></strong><a href="https://twitter.com/r3v3Ncl4W"><em>https://twitter.com/r3v3Ncl4W</em></a></p><p><strong><em>Instagram: </em></strong><a href="https://www.instagram.com/infosec.yash/"><em>https://www.instagram.com/infosec.yash/</em></a></p><p><strong><em>Blog page: </em></strong><a href="https://instagram.com/forensis_digitum"><em>https://instagram.com/forensis_digitum</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48afaa833c06" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your SOC Was Built for Breaches, Not Behaviour.]]></title>
            <link>https://yashgorasiya.medium.com/your-soc-was-built-for-breaches-not-behaviour-132e59cb502d?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/132e59cb502d</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[information-security]]></category>
            <category><![CDATA[soc]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 23 Feb 2026 09:01:01 GMT</pubDate>
            <atom:updated>2026-02-23T09:01:01.213Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bExCLFzd0MIB0nId.jpeg" /></figure><p>Nobody sent an alert when the model started lying.</p><p>There was no intrusion. No lateral movement. No exfiltration logged. The SIEM kept quiet. Endpoint detection reported nothing unusual. Uptime was green. The fraud detection model was processing thousands of transactions per hour, just as it always had. From the SOC’s perspective, everything was working exactly as intended.</p><p>Except the model had drifted. Quietly. Over weeks. The distribution of incoming transaction data had shifted as customer behaviour changed after a regional regulatory event, and the model, trained on patterns from a different economic moment, had recalibrated its confidence in ways no one had sanctioned. Approval rates in a specific demographic cohort had moved by eleven percentage points. Nobody noticed, because no control was designed to notice it.</p><p>This is the failure mode that does not have a ticket in your SIEM. It has no IP address. It triggers no rule. It leaves no forensic artifact that your SOC analyst can pull from a log at 2 AM.</p><p>This is where the gap lives. And it is material.</p><h3>How the SOC Was Built, and What It Was Built For</h3><p>The modern Security Operations Centre did not emerge from a concern about algorithmic outputs. It emerged from a concern about perimeter defence.</p><p>The architecture that most enterprise SOCs still run on, at its core, is event-based. Something happens: a login from an anomalous geography, a privilege escalation that violates a policy rule, a network packet that doesn’t match expected behaviour. That event creates a signal. The signal flows into a SIEM. Rules fire. Alerts queue. Analysts triage.</p><p>This model has a name in security engineering: detect-and-respond. It assumes that harm is introduced by something that crosses a boundary, leaves a trace, and creates a discrete, observable event. The Indicators of Compromise (IoC) doctrine runs through everything: IP addresses on threat intelligence feeds, hashes of known malicious files, network signatures matched against databases of adversary tooling. The entire intellectual infrastructure of the SOC, from its tooling to its staffing ratios to its escalation playbooks, was designed around this adversarial intrusion model.</p><p>That model is not wrong. Attackers do cross boundaries. They do leave traces. The detect-and-respond architecture has stopped real harm, at scale, for decades.</p><p>But it was never designed to monitor the behaviour of a probabilistic inference system deployed inside your own infrastructure. That is a categorically different problem.</p><p>When a SOC analyst opens a ticket, the question is: did something external attempt to harm us? When an AI model drifts, the question is: is something internal producing outputs that no longer reflect the intent we deployed it to fulfil? These are different questions. They require different telemetry, different thresholds, different ownership, and different escalation logic. Right now, in most organisations, the second question is owned by nobody with security authority.</p><h3>Event Signals and Behavioural Signals Are Not the Same Thing</h3><p>The distinction is worth being precise about, because conflating them is where most governance conversations go wrong.</p><p>An event signal is discrete. It has a timestamp, a source, a type. A failed authentication attempt at 3:47 AM is an event. A firewall block is an event. A hash matching a known ransomware signature is an event. These things either happened or they didn’t.</p><p>A behavioural signal is statistical. It describes a pattern over time. A model’s average confidence score across a class of inputs. The distribution of feature values arriving at an inference endpoint. The rate at which predictions cluster into particular output buckets. These signals don’t fire at 3:47 AM. They accumulate. They trend. They drift.</p><p>The table below illustrates the difference operationally:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/729/0*4-eY1v6sHkwH1rTi" /></figure><p><em>Source: Author’s framework, informed by NIST AI RMF 1.0 (January 2023), available at </em><a href="https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf"><em>https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf</em></a></p><p>This structural difference is not a nuance. It is the reason traditional SOC tooling, however mature and well-staffed, produces a blind spot at the exact point where AI systems fail most commonly.</p><p>An IBM analysis published in early 2025 put it plainly: traditional SOCs lack an adversarial framework for AI abuse, lack telemetry that distinguishes attack from drift, and lack the authority to intervene in AI pipelines. As a result, AI failures normalise as operational noise until business impact becomes visible, often too late for clean remediation. A model can be online, fully performant at an infrastructure level, and compliant with all access controls, while producing systematically distorted outputs. No amount of perimeter monitoring detects that condition.</p><p>That is not a gap in implementation. It is a gap in the architecture’s conceptual model.</p><h3>The Evidence Is Not Hypothetical</h3><p>There is a tendency in governance conversations to treat AI behavioural risk as a future concern, something to address once deployment matures. The evidence suggests we are already past that point.</p><p>According to the Stanford AI Index Report 2025, documented AI safety incidents surged from 149 in 2023 to 233 in 2024, a 56.4% increase in a single year. These are not theoretical harms. They are incidents captured across the AI Incident Database, which now catalogues over 1,000 real-world cases of AI systems causing or creating material risk of harm in deployed settings. The MIT AI Incident Tracker, which classifies incidents by causal factors and harm type, similarly documents more than 1,200 cases spanning physical harm, financial loss, human rights impacts, and property damage.</p><p>Critically, the distribution of causation matters here. Analysis of documented incidents suggests that approximately 67% stem from model errors, including drift, bias, and performance degradation, rather than adversarial attacks. Yet security budgets and SOC instrumentation are predominantly oriented toward the other 33%. This is not a rounding error. It represents a systematic misallocation of monitoring capacity relative to where the actual failure risk concentrates.</p><p>A 2024 study by the Mozilla Foundation found that 57% of companies using AI had no formal incident response plan in place, despite more than half of those same companies reporting at least one AI-related incident in the prior year. The detection gap compounds this: AI incidents take an average of 4.5 days to detect once they have occurred. In a fraud detection or credit risk context, 4.5 days of undetected model drift can translate into material harm at scale before anyone raises a ticket.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/744/0*8pjNWzrz8Vf8NZJd" /><figcaption>Source: MIT AI Incident Tracker</figcaption></figure><p>The NIST AI Risk Management Framework, in its 2023 core release and its July 2024 Generative AI Profile (NIST AI 600–1), explicitly calls for continuous monitoring of deployed AI systems for drift, bias, and performance degradation. The language in the framework is deliberate: it positions monitoring not as an MLOps housekeeping task but as a governance obligation, one that requires defined roles, thresholds, and escalation pathways. The framework is voluntary in the United States, but its influence is spreading into sector-specific regulatory postures globally.</p><p>In the UK, the Information Commissioner’s Office has issued guidance on accountability in automated decision-making under UK GDPR, including an expectation of ongoing monitoring for accuracy and fairness in deployed systems. The Financial Conduct Authority has been increasingly specific about model risk in its supervisory expectations for algorithmic decision-making in financial services, and the Bank of England’s Prudential Regulation Authority has referenced model monitoring as part of operational resilience frameworks for firms using AI in credit and risk functions. The EU AI Act, which introduces mandatory requirements for high-risk AI systems including those used in credit scoring, employment, and critical infrastructure, explicitly requires post-market monitoring plans with defined performance metrics and update processes.</p><p>The regulatory direction of travel is not ambiguous. What lags is the operational infrastructure to meet it.</p><h3>A Scenario: When the Model Drifts and the SOC Sees Nothing</h3><p>Consider a mid-sized retail bank operating in the UK and continental Europe. The bank deployed a machine learning model eighteen months ago to assist with personal loan origination decisions, flagging applications for manual review when the model’s confidence score falls below a defined threshold. The model was validated at deployment against the regulatory requirements of the time, documented in the AI risk register, and signed off by the second line.</p><p>The model’s training data reflected a specific economic period: pre-interest rate increases, stable employment in the bank’s core lending demographic, and consistent application patterns. Over the following fourteen months, several things changed. The macroeconomic environment shifted. Customer behaviour changed in ways that altered the statistical profile of incoming applications. A new broker channel, onboarded six months post-deployment, introduced application characteristics that were underrepresented in the training data.</p><p>None of these changes constituted a breach event. No external actor was involved. The model’s infrastructure was healthy. Log files showed normal inference throughput. The SIEM had nothing to report.</p><p>But the model’s decision boundary had quietly moved. The confidence threshold, set at deployment to reflect the model’s performance characteristics on its original data distribution, was no longer calibrated to the model’s actual confidence under the new distribution. The model was approving applications with lower true discriminatory power than the documentation suggested, and the pattern of who was being approved and declined had shifted in ways that correlated with protected characteristics.</p><p>This went undetected for eleven months. It was eventually surfaced not by any security or governance control, but by a customer service complaint pattern flagged by a compliance analyst reviewing a quarterly conduct report. The remediation involved model retraining, a retrospective review of affected decisions, regulatory notification, and a material provision for remediation costs.</p><p>No intrusion had occurred. The SOC had nothing to show for the period. The gap had not been in the perimeter; it had been in the lack of any control designed to observe the model’s behavioural outputs over time.</p><p><em>This scenario is illustrative.</em></p><h3>Where Visibility Drops Off: The AI Lifecycle and the SOC Boundary</h3><p>The SOC’s practical authority in most organisations runs from infrastructure upward: compute health, network telemetry, identity access management, endpoint protection, log aggregation. It stops, in most architectures, before it reaches the model layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*FbIHo1Hgv9bqo6vs" /><figcaption>Source: NIST AI RMF (Developed by the OECD).</figcaption></figure><p>The AI system lifecycle runs differently. A model passes through data curation, feature engineering, training, validation, staging, production deployment, and then ongoing inference. The risk profile across these stages is different from the risk profile the SOC was built to manage. The most consequential behavioural failures in AI systems tend to manifest in the inference phase, after deployment, when the real-world data distribution begins to diverge from the training distribution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ppC1uO_YIjihDjtW" /></figure><p>The structural gap is at inference time. That is where the model is actually operating on real data, making real decisions, in real conditions that may or may not resemble the conditions it was trained for. It is also the zone in which the SOC currently has the least native instrumentation.</p><p>MLOps teams have started building monitoring into this zone, using tools that track feature distributions, output distributions, and calibration metrics. But this monitoring sits inside ML engineering pipelines, not inside security governance. It is not connected to formal incident classification. It does not feed into escalation frameworks. It does not have a defined response procedure when a threshold is breached. And in many organisations, it is not happening at all.</p><h3>Operational Recommendations</h3><p>None of what follows is particularly radical. The uncomfortable part is not the recommendations themselves. It is that most organisations have had the conceptual tools to act on this for some time and have not.</p><p><strong>Behavioural observability as a security concern, not an engineering nicety.</strong> The SOC needs telemetry from the model layer. This means defining what constitutes a monitoring signal for each deployed AI system, output distribution shifts, confidence score trends, feature value anomalies, error rate changes across demographic slices, and routing those signals into a format the SOC can consume. This does not mean the SOC needs to become a data science team. It means the SOC needs a defined intake process for AI behavioural signals the way it has a defined intake process for network anomalies.</p><p><strong>Treating model drift as a formal incident class.</strong> Right now, most incident classification taxonomies include categories for data breach, ransomware, phishing, availability failure, and insider threat. They do not include a category for model behavioural degradation. This is not a philosophical omission. It means that when drift is eventually detected, there is no formal incident ticket, no ownership chain, no escalation path, and no remediation playbook. Creating that incident class, with defined severity levels, ownership, and escalation criteria, is a governance act that costs relatively little and closes a material gap. The EU AI Act’s Article 62 requirements for serious incident reporting to national competent authorities make this formalisation increasingly necessary for organisations deploying high-risk AI.</p><p><strong>Security integration into MLOps.</strong> The pattern that works for application security, shifting left and integrating security concerns into the development and deployment pipeline, applies to AI. This means security having visibility into model training data sourcing and integrity, model artefact handling, validation against defined behavioural benchmarks, and inference monitoring configuration. It does not require security teams to write model training code. It requires that the MLOps pipeline has security checkpoints with defined sign-off criteria, the same way a software deployment pipeline has code review and vulnerability scanning gates.</p><p><strong>Defining degradation thresholds before deployment, not after.</strong> The question of what constitutes unacceptable model performance should be answered at the governance stage, before the model goes live. What is the acceptable range for output distribution? At what confidence score degradation does the model require review? What demographic performance parity thresholds must be maintained? These are governance questions, and they need answers that are formally documented and technically enforced. Attempting to define degradation thresholds retrospectively, after a model has been running in production for a year, is structurally unreliable because people’s judgements are shaped by whatever the model has actually been doing.</p><p><strong>Escalation frameworks for behavioural anomalies.</strong> When a monitoring signal fires because a model’s output distribution has shifted by more than a defined tolerance, someone needs to own the response. That ownership should be defined, documented, and tested. It should specify who is notified, what assessment is required, what the remediation options are, and under what conditions the model is suspended from production. This is incident response logic applied to behavioural signals. It is achievable and it is absent from most current SOC playbooks.</p><h3>My POV</h3><p>I have sat in enough AI governance reviews to understand why this gap persists, and it is not because people are not intelligent or not trying.</p><p>What I have observed, repeatedly, is a jurisdiction problem that nobody wants to name directly. The SOC owns security events. The ML team owns model performance. The risk function owns the register. Legal owns regulatory exposure. And the space between these, where a model’s behavioural output constitutes an operational risk that is also a security concern and also a regulatory exposure and also a model performance issue, falls into a kind of organisational no man’s land where each function assumes another function is covering it.</p><p>The SOC will tell you, correctly, that model performance monitoring is not their remit. The ML team will tell you, also correctly, that they are monitoring performance metrics and the model is within acceptable accuracy bands. The risk function will point to the risk register, where the model is logged, validated at deployment, and rated amber. Nobody is wrong. And yet the model is drifting and nobody is catching it.</p><p>The insight that frameworks alone do not surface is this: the ownership gap at the behavioural layer is not resolved by assigning nominal accountability to a new function or creating a new working group. It is resolved by connecting existing monitoring signals to existing response infrastructure. The ML team already has some monitoring in their pipeline. The SOC already has escalation and response capability. The missing piece is an agreed interface between those two, a defined set of signals, thresholds, and escalation triggers that sits at the boundary of those two functions and creates an auditable chain of responsibility.</p><p>I have started calling this the observability handshake, not because it is a formal term, but because it captures what is actually needed. The ML engineer defines the signal. The security team defines the escalation response. Governance defines the threshold and the ownership. The interface between those three conversations is where organisations currently have the gap.</p><p>What concerns me is the regulatory acceleration timeline. The EU AI Act obligations for high-risk AI systems are becoming operational. The FCA and PRA are increasingly specific about model risk expectations. Organisations that do not have a formal behavioural monitoring and escalation pathway for deployed AI systems will find themselves in a position where they cannot demonstrate compliance with requirements that have actually existed in frameworks like NIST AI RMF since 2023.</p><p>The SOC did not build this gap. But it is positioned to be part of closing it. The question is whether security teams will advocate for the authority and instrumentation to do that, or whether they will continue to define their perimeter at the infrastructure layer and watch the behavioural risk accumulate just above it.</p><p>I know which option I would push for.</p><h3>Going Deeper on Substack</h3><p>If you’re responsible for AI risk management or trying to figure out how to go beyond quarterly reviews, the deeper analysis might be useful.</p><p><a href="https://cyberwithyash.substack.com/p/why-traditional-security-controls">Why Traditional Security Controls Can&#39;t See AI Failures</a></p><p>You can find it on Substack, where I publish monthly deep dives that go beyond what fits in the weekly LinkedIn format.</p><h3>Conclusion</h3><p>The SOC was not designed for this. That is not a criticism. It is an architectural fact.</p><p>The organisations that will navigate AI operational risk with least exposure are not the ones that buy a new tool or add a line to the RACI. They are the ones that build an honest interface between their ML engineering practice and their security operations capability, define what constitutes a behavioural incident before a model goes live, and treat drift as seriously as they treat an intrusion attempt.</p><p>Not because regulators will eventually require it. Although they will.</p><p>Because a model that is systematically producing the wrong outputs is causing harm. The absence of a breach signal does not change that. The SOC was built to stop harm. That mandate did not change when the harm started arriving without an IP address.</p><p><em>References and sources cited in this edition:</em></p><p><em>NIST AI Risk Management Framework 1.0 (January 2023): </em><a href="https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf"><em>https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf</em></a></p><p><em>NIST AI 600–1: Generative AI Profile (July 2024): </em><a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf"><em>https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf</em></a></p><p><em>Stanford AI Index Report 2025 (AI incident volume data):</em> <a href="https://aiindex.stanford.edu/report/">https://aiindex.stanford.edu/report/</a></p><p><em>AI Incident Database: </em><a href="https://incidentdatabase.ai/"><em>https://incidentdatabase.ai</em></a></p><p><em>MIT AI Incident Tracker: </em><a href="https://airisk.mit.edu/ai-incident-tracker"><em>https://airisk.mit.edu/ai-incident-tracker</em></a></p><p><em>Mozilla Foundation, AI Incident Response Study (2024): Referenced via AI Governance Lexicon / VerifyWise: </em><a href="https://verifywise.ai/lexicon/incident-management-for-ai-systems"><em>https://verifywise.ai/lexicon/incident-management-for-ai-systems</em></a></p><p><em>GLACIS AI Incident Response Playbook 2026 (incident detection timing, causal distribution data): </em><a href="https://www.glacis.io/guide-ai-incident-response"><em>https://www.glacis.io/guide-ai-incident-response</em></a></p><p><em>IBM Think Perspectives: “Why security operations must evolve for the AI era” (2025): </em><a href="https://www.ibm.com/think/perspectives/security-operations-evolve-ai-era"><em>https://www.ibm.com/think/perspectives/security-operations-evolve-ai-era</em></a></p><p><em>UK Information Commissioner’s Office — Guidance on Automated Decision-Making: </em><a href="https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/"><em>https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/</em></a></p><p><em>EU AI Act — Official Text (2024): </em><a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689"><em>https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=132e59cb502d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Rise of AI Red Teams: Testing the Machines Before the Machines Test Us]]></title>
            <link>https://yashgorasiya.medium.com/the-rise-of-ai-red-teams-testing-the-machines-before-the-machines-test-us-8cb10e4fc234?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/8cb10e4fc234</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[red-team]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[penetration-testing]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 10 Nov 2025 09:32:02 GMT</pubDate>
            <atom:updated>2025-11-10T09:32:02.969Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2VqDxGpqBTHwKEs-7eAhWg.png" /></figure><p>Welcome back to another blog…</p><p>If you’ve been following this series, you’ll know we’ve journeyed from the <strong>privacy paradox</strong> to <strong>AI governance</strong>, exploring how trust and technology are colliding in fascinating ways.</p><p>This week, let’s talk about something that sounds straight out of a spy thriller — <strong>AI Red Teams</strong>.</p><p>You’ve heard of penetration testers, right? The ethical hackers who try to break into systems before real attackers do?</p><p>Now imagine that concept — but for <strong>AI</strong>. Because today, the systems we need to secure aren’t just made of code — they <strong>think, reason, and learn</strong>.</p><p>And guess what? They can also <em>hallucinate, leak secrets, manipulate data</em>, and occasionally, <em>make things up with extreme confidence</em>. So who keeps these models honest?</p><p>Enter the <strong>AI Red Team</strong> — the unsung heroes (and sometimes professional troublemakers) who make sure your AI doesn’t go rogue before it hits production.</p><h3>What Exactly Is an AI Red Team?</h3><p>Let’s strip away the buzzwords.</p><p>An <strong>AI Red Team</strong> is a group of people (and sometimes, other AIs) who <strong>attack your AI system before someone else does</strong> — ethically, safely, and with the goal of making it stronger.</p><p>They don’t break your servers. They break your <strong>assumptions</strong>.</p><p>Here’s what they do in simple terms:</p><ul><li><strong>Try to make your AI say or do things it shouldn’t.</strong> Think of someone convincing your customer service chatbot to reveal pricing secrets or private customer info.</li><li><strong>Test for bias and manipulation.</strong> For example, can your hiring AI be tricked into preferring one gender or background just by changing word choices?</li><li><strong>Probe integrations.</strong> They check if your AI’s plug-ins, APIs, or data sources can be hijacked to cause chaos.</li><li><strong>Examine its “memory.”</strong> They try to make the model “remember” or regurgitate sensitive information from training data — like internal files or user records.</li></ul><p>Their job isn’t to embarrass the data science team — it’s to make sure that when the AI faces the real world, it doesn’t implode.</p><h3>Why Are AI Red Teams Suddenly a Big Deal?</h3><p>Two words: <strong>speed</strong> and <strong>scale</strong>.</p><p>AI systems are being rolled out faster than any tech in history. According to PwC’s 2025 UK AI Adoption Report, <strong>74% of UK enterprises</strong> have deployed at least one generative AI solution. Yet, less than <strong>27%</strong> have conducted structured AI security or risk assessments.</p><p>That gap — between <em>innovation</em> and <em>assurance</em> — is where AI red teams come in.</p><p>It’s not that companies don’t care about security. It’s that <strong>AI behaves differently from traditional software</strong>. You can’t just run a vulnerability scanner and call it a day.</p><p>Here’s why:</p><ul><li><strong>AI doesn’t follow instructions. It interprets them.</strong> That means an attacker doesn’t need to “hack” your system — they just need to <em>convince</em> it to misbehave.</li><li><strong>AI learns from data.</strong> So if your data has issues (biases, poisoned samples, outdated info), your model will faithfully reproduce them — or worse, amplify them.</li><li><strong>AI is connected to everything.</strong> Most enterprise models link to customer data, APIs, and cloud storage. So when AI breaks, it can break <em>big</em>.</li></ul><p>In the past, security was about closing ports. Now, it’s about closing <em>loopholes in language</em>.</p><h3>Traditional Pen Testing vs. AI Red Teaming</h3><p>Let’s make this easy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/797/1*4_5ovkhSjzGWCSe5CMvxnQ.png" /></figure><p>In short, traditional pentesting asks,</p><p>👉 <em>“Can I break in?”</em></p><p>AI red teaming asks,</p><p>👉 <em>“Can I make your system break itself?”</em></p><h3>Why This Shift Matters (and Why It’s Happening Now)</h3><p>Over the last year, we’ve seen <strong>AI models making it into critical workflows</strong> — healthcare summaries, legal document reviews, even banking chatbots.</p><p>The consequences of a single unsafe output can now be <strong>reputational, regulatory, and financial</strong>.</p><p>Let’s take a quick reality check:</p><ul><li><strong>In 2023</strong>, a major financial institution’s internal chatbot accidentally revealed confidential deal data in a private Slack conversation.</li><li><strong>In early 2024</strong>, researchers at Cornell and Google demonstrated that prompt injections could exfiltrate data from popular AI assistants, even when guardrails were “on.”</li><li><strong>UK’s National Cyber Security Centre (NCSC)</strong> released an advisory in 2024 warning that AI supply-chain vulnerabilities were now “a national concern.”</li></ul><p>AI red teaming became less of a “nice-to-have” and more of a <strong>compliance necessity</strong> — especially with the EU AI Act and the UK’s own emerging AI assurance frameworks.</p><p>The UK’s Department for Science, Innovation &amp; Technology (DSIT) even started working with private firms to define what “AI safety testing” should look like — a space where red teams play a central role.</p><h3>The Microsoft Example — When AI Testing Went Mainstream</h3><p>In 2023, Microsoft publicly revealed its internal <strong>AI Red Team</strong>, one of the first of its kind.</p><p>Their mission? To “think like adversaries, act like users, and discover unexpected behaviours before attackers do.”</p><p>What they found was eye-opening:</p><ul><li>Their own AI models could be tricked into revealing confidential information through indirect prompts.</li><li>Some models combined data from multiple sources in unpredictable ways, leading to accidental data leakage.</li><li>Updates meant to “improve performance” sometimes <strong>reintroduced old vulnerabilities</strong>, a concept known as <strong>model regression</strong>.</li></ul><p>The takeaway? AI safety isn’t a <em>project</em>. It’s a <strong>process</strong>. And that process must include someone deliberately trying to outsmart the model — before the bad guys do.</p><h3>The Rise of the Red Team Economy</h3><p>Globally, we’re now seeing an emerging industry around AI testing.</p><p>Companies like <strong>Anthropic</strong>, <strong>OpenAI</strong>, and <strong>Google DeepMind</strong> have formal red-teaming programs that involve experts from outside their walls.</p><p>Startups in the UK — like <strong>Lakera</strong>, <strong>Robust Intelligence</strong>, and <strong>HiddenLayer</strong>…are building tools to automate AI vulnerability discovery and resilience testing.</p><p>The NCSC and GCHQ-backed <strong>AI Security Conference</strong> in London this year even dedicated an entire track to “AI Adversarial Testing and Red Teaming.” That’s how mainstream this has become.</p><p>And just to put numbers behind it: according to <strong>Cyber UK 2025 reports</strong>, spending on AI assurance and adversarial testing in the UK is expected to grow <strong>8x by 2027</strong>.</p><h3>The People Behind the Tests</h3><p>Here’s the part many organisations underestimate — <strong>AI red teams aren’t just hackers with hoodies</strong>.</p><p>A typical team includes:</p><ul><li><strong>AI engineers</strong> (who understand how models think),</li><li><strong>Security researchers</strong> (who know how attackers think),</li><li><strong>Data scientists</strong> (who spot biases and weird data behaviour),</li><li><strong>Legal &amp; compliance folks</strong> (who make sure testing doesn’t cross the line), and</li><li><strong>Psychologists or linguists</strong> occasionally to test manipulation and bias in communication-based models.</li></ul><p>Because AI failure isn’t always technical — sometimes, it’s <strong>psychological</strong>.</p><h3>Why the UK Is Taking It Seriously</h3><p>The UK has been quietly leading the global conversation on <strong>AI assurance</strong>.</p><p>In 2024, the <strong>NCSC</strong> released a <strong>Guideline for Secure AI System Development</strong>, urging companies to:</p><ul><li>Conduct <strong>red-team testing</strong> before deployment,</li><li>Maintain model and data provenance,</li><li>And monitor real-world behaviour continuously.</li></ul><p>Meanwhile, <strong>DSIT</strong>’s <em>AI Assurance Roadmap</em> highlighted red teaming as a <strong>critical pillar</strong> of responsible AI adoption — right next to governance and transparency.</p><p>For UK enterprises — especially those in financial services, healthcare, and government — red teaming is no longer just a tech buzzword. It’s an <strong>expectation</strong>.</p><h3>The Anatomy of an AI Red Team: Tools, Tactics, Workflow (Practical &amp; Technical)</h3><p>This is the section you’ll hand to your engineering leads and say: “Do this next quarter.”</p><h3>1) Start with the basics: scope, safety, and legal sign-off</h3><p>Before you run anything, run the paperwork.</p><ul><li><strong>Rules of Engagement (RoE)</strong> — document who may test, what targets (endpoints, model versions), exact allowed techniques, time windows, and emergency stop criteria. This must be signed by security, legal, privacy, and the business owner.</li><li><strong>Data safety boundaries</strong> — define what data may be used for testing. Use synthetic or anonymised data wherever possible. If you must use production-like data, apply strict controls, short retention, and privacy review.</li><li><strong>Kill switch &amp; incident plan</strong> — know how to immediately stop tests and revert model endpoints. Define notification timelines and decision authority.</li></ul><p>This is not bureaucracy. This is insurance. Without it, you risk legal trouble and accidental exposure.</p><h3>2) Build your inventory first — AI SBOM (Software Bill of Materials)</h3><p>You cannot test what you cannot name.</p><p>Create an <strong>AI SBOM</strong> for your product. At minimum, track:</p><ul><li>Model names and exact versions (base model + any fine-tune/version hash)</li><li>Training/fine-tuning datasets (names, owners, dates)</li><li>Embedding models and vector DB versions</li><li>Plugins, connectors, and third-party APIs + their access scopes</li><li>Serving infra (containers, cloud regions, ingress points)</li><li>Model card/datasheet links, and vendor audit status</li></ul><p>Practical tip: store the SBOM in a living repo (Git-backed) and require updates via CI when models or dependencies change.</p><h3>3) Threat modelling — make it concrete</h3><p>A good red team starts with a crisp threat model. Ask:</p><ul><li><strong>Who</strong> are we defending against? (insider, script-kiddy, nation-state, curious researcher)</li><li><strong>What</strong> do they want? (PII, intellectual property, sabotage, biased decision outcomes)</li><li><strong>Where</strong> can they act? (public UI, API, partner plugin, misconfigured S3)</li><li><strong>How</strong> could they accomplish it? (prompt injection, gradual extraction, model poisoning, exploit infra)</li></ul><p>Document attacker capabilities (API access, auth, rate limits) and map attack paths. This keeps tests focused and measurable.</p><h3>4) Test design — black-box, grey-box, white-box</h3><p>Design 3 classes of tests:</p><ul><li><strong>Black-box</strong> — tester uses the public UI/API, no internal access. This mimics an external attacker and is safe for production if scoped. Use for prompt injection, output-sanity, and monitoring checks.</li><li><strong>Grey-box</strong> — tester has some internal knowledge (model card, config, sample logs). Useful for targeted privacy checks and chained-behaviour tests.</li><li><strong>White-box</strong> — full access to model weights, pipelines, and training data. Reserved for lab environments; used to test poisoning, gradient-based attacks, extraction feasibility, and explainability probes.</li></ul><p>Choose the class per risk tolerance. For customer-facing models, black-box + grey-box is often the balance; white-box for internal R&amp;D or offline experiments.</p><h3>5) Practical testing categories — what you actually run</h3><p>Organise tests into repeatable suites. Below are practical categories with their intent and safe methods.</p><p><strong>A. Prompt-safety &amp; jailbreak suite</strong> Purpose: measure how often the model ignores system instructions and emits unsafe or disallowed content. Methods (safe): curated prompt corpora, paraphrase families, and chained-context sequences. Track pass/fail and severity.</p><p><strong>B. Context leakage &amp; membership checks</strong> Purpose: ensure the model doesn’t leak training-set fragments or PII. Methods (safe): use synthetic “canary” phrases included in non-sensitive test corpora to see if the model regurgitates them; check for verbatim reproductions of known training fragments.</p><p><strong>C. Model-extraction estimation</strong> Purpose: estimate how many queries an adversary would need to approximate model behavior. Methods (safe): measure response diversity vs. query volume; test with rate-limited small sample sets to model extraction cost. (Do not attempt full reconstruction unless sanctioned in an isolated lab.)</p><p><strong>D. Supply-chain verification</strong> Purpose: check integrity and provenance of third-party weights/plugins. Methods: verify cryptographic hashes, require signed models, check vendor SBOMs, and scan containers for CVEs.</p><p><strong>E. Operational resilience tests</strong> Purpose: validate monitoring, alerting, rate limits, and rollback behavior. Methods: low-and-slow simulated extraction, burst traffic tests (within agreed safe limits), verify that alerts map to incidents and that runbooks trigger.</p><p><strong>F. Bias &amp; fairness probes</strong> Purpose: surface model biases or discriminatory outputs. Methods: behaviour-driven tests (CheckList-style), test suites across demographic dimensions, and scenario-based evaluation.</p><h3>6) Toolbelt — what to use and why</h3><p>Here’s the practical, field-tested toolset with short notes. (Use these as building blocks — many teams combine them.)</p><p><strong>Prompt fuzzing &amp; adversarial NLP</strong></p><ul><li><strong>TextAttack</strong> — create adversarial paraphrases and fuzzers; script attack campaigns.</li><li><strong>OpenAttack</strong> — an alternative suite for NLP attacks and pipelines.</li></ul><p><strong>Robustness &amp; privacy/membership</strong></p><ul><li><strong>IBM Adversarial Robustness Toolbox (ART)</strong> — membership inference, poisoning simulations, and defences.</li><li><strong>SecML</strong> — for lab experiments (poisoning/evasion prototyping).</li></ul><p><strong>Behavioural testing &amp; regression</strong></p><ul><li><strong>CheckList (Microsoft)</strong> — build test catalogs and behaviour suites for capabilities/failure modes.</li><li><strong>Custom corpora</strong> — curate domain-specific prompts; maintain as code (tests-as-code).</li></ul><p><strong>Provenance &amp; observability</strong></p><ul><li><strong>MLflow / Weights &amp; Biases</strong> — track model versions, artifacts, and experiments.</li><li><strong>WhyLabs / Evidently</strong> — runtime monitoring for distribution drift and data anomalies.</li></ul><p><strong>Infra &amp; dependency scanning</strong></p><ul><li><strong>Snyk / Trivy / Dependabot</strong> — scan containers and libraries for CVEs.</li><li><strong>Clair</strong> — container image analysis for vulnerabilities.</li></ul><p><strong>Serving + control platforms</strong></p><ul><li><strong>Seldon Core / BentoML / KFServing</strong> — for serving with metrics, canarying and hooks for logging.</li><li><strong>Prometheus + Grafana</strong> — telemetry and anomaly dashboards.</li></ul><p><strong>Orchestration &amp; harnesses</strong></p><ul><li>Build small Python harnesses that bulk-submit prompts, collect outputs, tag model metadata, and push to a central ELK or MLflow for analysis.</li></ul><p>PS. I am not affiliated with any of the above vendors. Try at your convenience.</p><h3>7) Test automation &amp; regression — make it repeatable</h3><p>Automate behavioural regression tests just like unit tests:</p><ul><li>Store prompt corpora as code and run them in CI on model updates.</li><li>Fail the build if critical safety tests degrade.</li><li>Version the corpora and tag passing thresholds per model release.</li></ul><p>Automation ensures you don’t ship regressions accidentally when a vendor updates a weight or you fine-tune.</p><h3>8) Evidence &amp; reporting — what a board wants to see</h3><p>A good red-team report is not an academic paper. Deliverables should include:</p><ul><li><strong>Executive summary</strong> — one page: scope, top 3 risks, business impact, remediation priorities.</li><li><strong>Findings catalog</strong> — reproducible test cases, inputs, outputs, model metadata, timestamps.</li><li><strong>Risk rating &amp; owner map</strong> — severity and who will fix it (ML, infra, vendor).</li><li><strong>Remediation roadmap</strong> — quick wins, medium-term fixes, long-term changes.</li><li><strong>KPIs</strong> — jailbreak rate, MTTD (mean time to detect), MTTR (mean time to remediate), % of models with SBOM.</li></ul><p>Boards and auditors love traceability. Make your evidence stash searchable and auditable.</p><h3>9) Operationalise fixes — from red-team to “release-gate”</h3><p>Don’t let findings gather dust. Embed fixes into the product lifecycle:</p><ul><li><strong>Release gates</strong> — require passing behavioral regression &amp; SBOM checks before promotion to prod.</li><li><strong>Canary + staged rollouts</strong> — new model versions go to a small, monitored cohort.</li><li><strong>Monitoring &amp; alerts</strong> — map specific red-team indicators to production alerts (e.g., spike in “hallucination” tokens).</li><li><strong>Vendor SLAs &amp; contracts</strong> — require model signatures, notice of retrainings, and rights to audit.</li></ul><p>The flow you want: red team finds → product prioritises → devs fix → tests validate → deploy.</p><h3>10) Sample safe test (high level) — membership canary (how it helps)</h3><p>A quick conceptual test you can run with safety:</p><ul><li>Insert a unique but harmless “canary string” into a training/test corpus in a lab (e.g., “CANARY-2025-XY”).</li><li>Later, query the model with prompts likely to elicit related content.</li><li>If the model returns the canary verbatim, you have evidence of memorisation/regurgitation.</li><li>Use this to measure privacy leakage risk and calibrate redaction or differential privacy needs.</li></ul><p>This is a safe, ethical way to quantify a leakage problem without exposing real PII.</p><h3>11) KPIs to run and report monthly</h3><p>Make red-team work visible with measurable trends:</p><ul><li><strong>Jailbreak rate per 1,000 queries</strong> (downward trend target)</li><li><strong>Canary leakage occurrences</strong> (count &amp; severity)</li><li><strong>Time to detect extraction attempts (MTTD)</strong></li><li><strong>Time to remediate red-team findings (MTTR)</strong></li><li><strong>% of models with full SBOM &amp; signed weights</strong></li><li><strong>Number of vendor components with audit rights</strong></li></ul><p>Numbers convert effort into risk metrics the business understands.</p><h3>12) People &amp; culture — the non-technical glue</h3><p>Finally, red teams need culture to be effective:</p><ul><li>Make red-team findings <em>non-punitive</em>. It’s about product quality, not blame.</li><li>Rotate engineers into the red team occasionally to build empathy and prevent stovepipes.</li><li>Publicise wins (reduced jailbreak rate, faster patch cycles) to get continued investment.</li></ul><p>Security is a team sport. The most defensible AI programs are the ones where product, ML, infra and legal all sleep easier because they built testing into how they work.</p><h3>Lessons from the Field</h3><p>Case studies help us move from abstract warnings to concrete fixes. Below are anonymised, realistic summaries based on public incidents, industry reporting, and red-team outcomes. I’ve kept them practical and focused on lessons — not on finger-pointing.</p><h3>Case Study 1 — The Meeting Summary that Wasn’t Private</h3><p><strong>What happened:</strong> A large professional services firm deployed an internal “meeting summariser” using a hosted LLM. The model was configured to pull meeting notes, recordings, and shared slides to create concise summaries for employees. During a red-team style review (triggered by an internal audit), engineers found that some summaries occasionally included verbatim snippets from confidential client presentations.</p><p><strong>Root cause:</strong> The model’s context window included verbatim content from linked slide decks; the post-processing layer didn’t redact or filter proprietary phrases. In addition, the vendor’s logging retained prompt outputs for longer than contractually allowed.</p><p><strong>Impact:</strong> Potential exposure of client IP and a breach of contractual confidentiality clauses. The firm avoided public disclosure because they caught it internally, but the ICO likely would have considered it reportable.</p><p><strong>Fix:</strong></p><ul><li>Implement deterministic post-processing filters to remove direct verbatim sequences above a token threshold.</li><li>Limit retrieval context to headline-level metadata, not full slide text.</li><li>Contractually require vendor logging retention limits and ability to purge.</li><li>Add canary phrases to test outputs regularly.</li></ul><p><strong>Lesson:</strong> Guardrails matter. If your model pulls content, ensure you control how much of it is visible to generation, and verify vendor behaviour matches contract.</p><h3>Case Study 2 — The Plugin That Knew Too Much</h3><p><strong>What happened:</strong> A fintech company integrated a third-party “payments helper” plugin into its customer support assistant. The plugin was supposed to fetch payment statuses. During a routine red-team probe, testers discovered that chaining certain benign queries with the plugin returned aggregated views of invoices across customers — an over-privileged behaviour.</p><p><strong>Root cause:</strong> The plugin was authorized with overly broad OAuth scopes. The risk review process for third-party plugins was immature: procurement had not verified least privilege design.</p><p><strong>Impact:</strong> Exposure of aggregated financial data — significant reputational and regulatory risk for a regulated sector.</p><p><strong>Fix:</strong></p><ul><li>Revoke and reissue tokens with least-privilege scopes.</li><li>Implement a plugin vetting checklist (access scopes, vendor SBOM, CVE history).</li><li>Introduce runtime checks that map plugin queries to the calling user’s tenant or identity.</li><li>Add automated unit tests that simulate cross-tenant scenarios.</li></ul><p><strong>Lesson:</strong> Third-party integrations are a primary attack surface. Treat plugins like code you deploy: vet, scope, and monitor.</p><h3>Case Study 3 — The Supply-Chain Poisoning Near-Miss</h3><p><strong>What happened:</strong> A mid-sized SaaS company used an open-source embedding model from a public repo for semantic search. A security researcher discovered malicious content (a tiny hidden script in a model artifact) in a community package. The company’s automated CI pulled the artifact automatically during a weekend build. Fortunately, the pipeline’s pre-deploy checks detected a checksum mismatch (because the company had enforced model signing). The incident was contained.</p><p><strong>Root cause:</strong> Reliance on a public repo without continuous integrity checks. The attacker targeted the community package supply chain.</p><p><strong>Impact:</strong> Potential for silent exfiltration or corrupted embeddings used across multiple customer tenants. Containment avoided a breach.</p><p><strong>Fix:</strong></p><ul><li>Require signed models and enforce checksums in CI.</li><li>Maintain a curated internal registry of vetted model artifacts.</li><li>Implement SBOMs that trace model origin and signature.</li><li>Add periodic artifact integrity scans.</li></ul><p><strong>Lesson:</strong> Model provenance and artifact signing are not optional. Treat models like software packages: sign, verify, and lock your supply.</p><h3>Case Study 4 — The Hallucination That Cost Money</h3><p><strong>What happened:</strong> An e-commerce company used a model to auto-generate product descriptions and suggested shipping instructions. A model update intended to improve fluency unintentionally increased hallucination propensity: the model began inventing non-existent stock-keeping units (SKUs) and suggesting deprecated discount codes in customer-facing outputs. The issue wasn’t obvious until customers started complaining and support tickets spiked.</p><p><strong>Root cause:</strong> Insufficient pre-deploy behavioural regression tests that measure hallucination rates. The vendor model update changed output distribution without backward compatibility testing.</p><p><strong>Impact:</strong> Customer trust erosion, lost sales, and a three-week rollback until the older model was reinstated.</p><p><strong>Fix:</strong></p><ul><li>Add behavioural regression suites to CI that include hallucination detection tests (e.g., fact-check against inventory).</li><li>Implement staged canary rollouts and automated rollback triggers on metric deviations.</li><li>Contractual requirement: vendors must notify of significant model updates and provide changelogs.</li></ul><p><strong>Lesson:</strong> Behavioural regression is as important as functional regression. Always test for changes in output distribution.</p><h3>The regulatory and policy tie-in — what the UK and global frameworks expect</h3><p>Understanding the rules helps you prioritize actions that matter. Below I map the red-team theme to the key regulatory expectations, particularly focusing on the UK with global context.</p><h3>The EU AI Act (short, practical take)</h3><p>The EU AI Act classifies systems by risk. High-risk systems (e.g., recruitment tools, credit scoring, critical infrastructure) require:</p><ul><li>Conformity assessments,</li><li>Detailed technical documentation,</li><li>Ongoing monitoring and testing,</li><li>Transparency measures, and</li><li>Human oversight mechanisms.</li></ul><p><strong>Relevance to red-teaming:</strong> If your model falls under “high-risk,” you will need documented adversarial testing as evidence in conformity assessments. Model provenance and SBOMs help fulfil documentation requirements.</p><h3>UK landscape — ICO, NCSC, DSIT</h3><ul><li><strong>ICO (Information Commissioner’s Office)</strong>: focuses on data protection. If red-team tests expose or show plausible paths to PII leakage, you must treat it as a DPIA concern and possibly a reportable breach. ICO expects data controllers to demonstrate due diligence and reasonable safeguards.</li><li><strong>NCSC (National Cyber Security Centre)</strong>: issues practical secure development guidance; emphasises adversarial testing and supply-chain integrity. NCSC guidance is operationally useful for security teams.</li><li><strong>DSIT / AI Assurance Roadmap</strong>: the UK’s stance is principle-led; institutions should show evidence of testing, risk management, and proportionate governance. The AI Safety Institute pilot programs also expect red-team participation for higher-risk systems.</li></ul><p><strong>Practical takeaway for UK firms:</strong> maintain red-team reports, SBOMs, and monitoring logs as compliance evidence. Expect to present these to auditors or regulators.</p><h3>US and other jurisdictions</h3><ul><li><strong>US Executive Orders and sectoral guidances</strong> encourage adversarial testing and vendor transparency but rely more on sector regulators (e.g., FTC, SEC).</li><li><strong>China</strong> imposes tighter pre-approval and content rules for large models — a different compliance model, more prescriptive.</li></ul><p><strong>Global view:</strong> compliance is patchy but trending: regulators want demonstrable assurance, and red-teaming is becoming a practical way to provide it.</p><h3>My POV: The Future of Red Teams Isn’t Just About Breaking Things — It’s About Understanding Them</h3><p>In my view, <strong>AI red teaming will become the new digital due diligence</strong>. Before any enterprise pushes a model into production, red teaming will be as normal as pen testing an app. You’ll have entire teams whose job is to find ways AI could mislead, discriminate, or leak — and fix those issues before customers, regulators, or journalists do.</p><h3>AI Red Teams Will Blur Boundaries</h3><p>Over the next few years, I see a shift where red teams and data scientists won’t sit in silos anymore. They’ll work together in cycles — train, test, break, rebuild. A continuous loop of resilience.</p><p>Imagine this:</p><ul><li>Your security team uses AI to simulate thousands of phishing attempts against your chatbot.</li><li>Your data scientists tweak the model to resist those attacks.</li><li>Your compliance lead maps that improvement directly to regulatory requirements.</li></ul><p>That’s what mature AI assurance looks like — a constant feedback loop between humans, models, and governance.</p><h3>Governance Will Need a Personality Upgrade</h3><p>Governance frameworks today often feel like checklists — rigid and slow. But the future of AI security demands something more adaptive. Governance needs to evolve from “policing” to <strong>“coaching.”</strong> Instead of just saying “no” to risky deployments, AI governance teams should be equipped to say, “Here’s how to make it safe enough to say yes.”</p><p>The EU AI Act, UK’s AI Safety Institute, and US NIST frameworks are laying early groundwork, but the next leap will come from enterprises themselves — the ones that bake responsible AI into their DNA, not their paperwork.</p><h3>The UK’s Sweet Spot</h3><p>The UK actually sits in a unique position here. With the <strong>National Cyber Security Centre (NCSC)</strong> already championing AI safety and <strong>London’s vibrant AI startup ecosystem</strong>, the region can pioneer <em>practical</em> AI red teaming — frameworks that combine innovation with accountability. The challenge, however, is scale: how do mid-sized enterprises join this journey without the budgets of Big Tech? That’s where open-source tools, shared learning platforms, and collective testing efforts could make all the difference.</p><h3>From “Security for AI” to “AI for Security”</h3><p>This transition fascinates me. We started this series months ago talking about how to <em>protect AI</em> — today, we’re already seeing AI being used to <em>protect everything else</em>.</p><p>AI red teaming embodies that shift perfectly. It’s no longer about defending machines from humans — it’s about teaching machines to defend themselves, ethically and intelligently.</p><p>And maybe that’s the real takeaway here: <strong>AI red teaming isn’t just a security practice. It’s a mirror.</strong> It reflects our biases, our assumptions, and our blind spots back at us — forcing us to question whether our AI systems are learning the best parts of humanity or the worst.</p><h3>The Final Sip</h3><p>If you’ve read this far, first — thank you.</p><p>Red teaming AI systems isn’t just about keeping up with regulators or winning trust. It’s about accountability. The kind that makes technology not only more secure but also more human.</p><p>Because the real test of AI isn’t whether it passes our red teams — it’s whether it passes the test of time.</p><p><strong>Thank you for reading this blog.</strong> If this sparked ideas or questions, drop them in the comments or DM me. I would love to hear how <em>you</em> think AI assurance will evolve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ijkBiwWhoGJ8D0j30NVCaQ.png" /></figure><p>☕ Until next week — keep your models sharp, your prompts ethical, and your curiosity alive.</p><p>Please don’t forget to leave a review.</p><p>If you found this helpful, subscribe to stay in the loop for future editions.</p><p>Together, let’s build a safer and more secure digital world! 💪🤗👨💻</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8cb10e4fc234" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI’s Shadow Supply Chain: The Hidden Risks Behind Every Prompt]]></title>
            <link>https://yashgorasiya.medium.com/ais-shadow-supply-chain-the-hidden-risks-behind-every-prompt-71fcfdf8038b?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/71fcfdf8038b</guid>
            <category><![CDATA[software-supply-chain]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[risk-management]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Tue, 28 Oct 2025 08:02:00 GMT</pubDate>
            <atom:updated>2025-10-28T08:02:00.157Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MH3OY-daGKXufnn6Dghhew.png" /></figure><p>Welcome to another blog on AI Security. I hope everyone has had a refreshing festive season. 🎉</p><p>The lights, the laughter, the break from dashboards — all much needed. And now that we’re back on track, let’s talk about something less sparkly but far more important: the shadows hiding inside our AI systems.</p><p>Because while most of us were busy celebrating, the algorithms never stopped working — or, in some cases, <em>depending</em> on things we don’t even know exist.</p><h3>The Hidden Machine Behind the Magic</h3><p>Imagine you walk into a coffee shop, order a cappuccino, and the barista proudly says,</p><blockquote><em>“Everything here is handcrafted.”</em></blockquote><p>Then you glance behind the counter and see ten machines grinding, steaming, frothing — each made by a different supplier, each with its own quirks.</p><p>That, in essence, is modern AI.</p><p>Every slick AI assistant, predictive engine, or chatbot that enterprises deploy depends on dozens — sometimes hundreds — of unseen components: open-source model weights, third-party APIs, data pipelines, vector databases, and plugin integrations. Together they form what experts are now calling the <strong>AI supply chain</strong>.</p><p>The “shadow” part comes from what we <em>don’t</em> see: outdated libraries quietly imported from GitHub, experimental datasets scraped from who-knows-where, or micro-vendors providing model hosting services without any formal security audit.</p><p>According to Gartner’s 2025 State of AI Infrastructure Report, over <strong>80 percent</strong> of enterprise AI systems rely on at least one open-source or third-party model component. Only <strong>22 percent</strong> of those components undergo a full security review before deployment.</p><p>If that statistic made you exhale loudly, good. You’re paying attention.</p><h3>When the Chain Becomes a Maze</h3><p>Let’s take a typical corporate AI project — say, a “smart document summariser” built for internal reports.</p><p>It might use:</p><ul><li><strong>A fine-tuned open-source LLM</strong> from Hugging Face,</li><li><strong>A vector database</strong> from a small cloud vendor,</li><li><strong>A content-filtering API</strong> from another startup,</li><li><strong>An orchestration layer</strong> connecting it all together,</li><li>And, of course, <strong>a pre-trained embedding model</strong> someone found on a developer forum.</li></ul><p>Each piece may function perfectly on its own. But connect them all, and you have an ecosystem whose weakest link defines your entire security posture.</p><p>If any one vendor suffers a breach, if any dependency is poisoned, if any dataset contains sensitive information, your AI inherits the problem instantly — often invisibly.</p><p>That’s why the National Cyber Security Centre (NCSC) warned this year that <em>“AI systems inherit the trust boundaries of every model and service they integrate.”</em></p><p>In plain English: when you adopt someone else’s model, you adopt their mistakes.</p><h3>How We Got Here</h3><p>The irony is that this wasn’t negligence — it was enthusiasm. The open-source AI boom made building complex systems feel easy. Why spend months training a model when you can download one that’s “good enough” in minutes?</p><p>Corporate teams embraced modular AI the same way children embrace LEGO: mix pieces, build fast, admire the result. Only now we’re discovering that some bricks are made of unstable material.</p><p>Take, for example, the 2025 <strong>Hugging Face repository poisoning incident</strong>. A researcher discovered that several public model files contained hidden payloads capable of exfiltrating credentials when loaded. Thousands of downloads later, the vulnerability was disclosed.</p><p>No catastrophic breach followed — this time. But it highlighted how a compromised model on a public repository could travel through the supply chain at lightning speed, quietly embedding itself in commercial systems.</p><h3>The Ghost of SolarWinds</h3><p>If this sounds familiar, it should. Remember the 2020 SolarWinds attack, where hackers inserted malicious code into a software update used by 18,000 organisations, including multiple governments?</p><p>That was the classic <em>software</em> supply chain failure. Now we’re seeing its AI sequel.</p><p>The difference is subtle but significant: traditional software delivers <strong>functions</strong>; AI systems deliver <strong>behaviours</strong>. When a library is compromised, you can patch the code. When a model is compromised, you might have to retrain it — assuming you can even trace where the taint came from.</p><p>AI doesn’t just execute logic; it interprets, learns, and generates. That makes its vulnerabilities creative — and far harder to contain.</p><h3>What Exactly Is in the Chain?</h3><p>Let’s unpack the main layers of an AI supply chain — because knowing the anatomy helps you find the infection.</p><ol><li><strong>Data Supply</strong> — Every dataset used for pre-training or fine-tuning: text corpora, images, transaction records. The risk? Poisoned or biased data.</li><li><strong>Model Supply</strong> — Base models, fine-tunes, and embeddings. Many come from open hubs with minimal provenance checks.</li><li><strong>Infrastructure Supply</strong> — Cloud hosting, APIs, orchestration tools. Breaches here can expose entire model pipelines.</li><li><strong>Plugin / Extension Supply</strong> — External integrations (like productivity-suite connectors) that expand capability — and attack surface.</li><li><strong>Human Supply</strong> — Developers, prompt engineers, data labellers. Insider errors or malicious actions remain perennial risks.</li></ol><p>Each layer depends on trust. But trust, as every security professional knows, is not a control — it’s a vulnerability politely waiting for validation.</p><h3>The Invisible Intermediaries</h3><p>Here’s where things get truly opaque.</p><p>Most enterprises don’t deal directly with model creators; they use intermediaries — cloud providers, integrators, or vendors who bundle multiple AI services into a neat package.</p><p>So when an incident occurs, tracing responsibility becomes a corporate version of <em>“Who, Me?”</em></p><p>Regulators ask, “Whose model was it?” Vendors point to open source. Open-source maintainers point to contributors. Contributors point to datasets scraped from the web. And somewhere in that circle sits your company’s name on the ICO’s inquiry letter.</p><p>This fragmentation means AI risk management can’t rely on one-time vendor assessments. It requires continuous visibility — provenance tracking, integrity checks, and dependency scanning.</p><p>Unfortunately, as of 2025, fewer than one-third of large UK firms have any formal process to track model provenance end-to-end (source: Deloitte AI Risk Survey 2025).</p><h3>Why It Matters Now</h3><p>Because the supply chain is getting longer by the week.</p><p>Every major LLM vendor is launching plugin ecosystems, partner marketplaces, and model-as-a-service offerings. Each new integration promises productivity — and imports a new variable of trust.</p><p>As NCSC’s <em>Secure AI Framework</em> puts it:</p><blockquote><em>“Complexity expands faster than assurance.”</em></blockquote><p>That’s the quiet danger. The very modularity that makes AI exciting — plug, play, scale — also makes it porous.</p><p>The question isn’t <em>if</em> a supply-chain compromise will occur. It’s <em>how prepared</em> organisations will be when it does.</p><h3>The Real-World Wake-Up Calls</h3><p>We’ve seen warning shots already — small incidents that hint at much bigger storms ahead.</p><p>Earlier this year, researchers found that several <strong>open-weight AI models</strong> hosted on public repositories had been tampered with to include malicious instructions. The compromise didn’t target enterprises directly; it targeted <em>their dependencies</em>.</p><p>One affected model was downloaded by a London-based SaaS firm building a “smart meeting assistant.” The developers thought they were using a safe version of a speech-to-text model. What they actually got was one that quietly sent usage logs to an external server. The incident never made headlines — the company caught it early — but it exposed how fragile the trust chain really is.</p><p>A similar scare came from <strong>Hugging Face</strong>, which issued an urgent bulletin in mid-2025 urging users to verify model integrity after discovering credential exfiltration payloads embedded in certain repositories.</p><p>If this sounds niche, remember: Hugging Face hosts over <strong>500,000</strong> models used across academia, startups, and Fortune 500s. One poisoned repository isn’t a blip — it’s a potential contagion.</p><p>The National Cyber Security Centre (NCSC) quietly advised UK businesses to “treat third-party AI models as untrusted inputs unless verified.” It’s the digital equivalent of washing your vegetables before cooking — something obvious that still isn’t common practice.</p><h3>Echoes of the Past: From Log4j to LLMs</h3><p>To understand how this can spiral, rewind to 2021’s <strong>Log4j</strong> vulnerability — a single open-source logging library flaw that shook the software world. One component, used in thousands of applications, opened doors for attackers globally.</p><p>The AI world is heading toward the same cliff. Except this time, the component isn’t a few lines of code — it’s an entire reasoning engine that can interpret instructions.</p><p>When a model dependency is compromised, the effects aren’t just technical; they’re behavioural. It might misclassify data, leak sensitive prompts, or generate subtly biased results that warp downstream analytics. That’s harder to detect and even harder to explain.</p><p>SolarWinds taught us that trust is fragile. Log4j taught us that ubiquity is risk. AI will teach us that <strong>intelligence multiplies both.</strong></p><h3>Regulatory Reactions: Trying to Shine Light on the Shadows</h3><p>Regulators are beginning to notice.</p><p>The <strong>EU AI Act</strong>, now in effect, includes language requiring providers of “high-risk” systems to maintain detailed technical documentation of their training data, model lineage, and third-party dependencies. In theory, that creates accountability. In practice, it creates paperwork.</p><p>Many vendors have already started offering “model cards” — neat PDFs listing datasets, architectures, and fine-tune details. They look reassuring, but often omit what matters most: the open libraries or plug-ins that bridge one model to another.</p><p>The <strong>US</strong> is pushing a different angle through its <em>AI Safety Institute Consortium</em> — encouraging industry to share incident data voluntarily. China continues its tightly controlled model-approval regime, where every large-scale model must pass a government audit before release.</p><p>And the <strong>UK</strong>, ever the pragmatist, relies on guidance rather than punishment. The NCSC’s <em>Secure AI System Development Framework</em> and the Department for Science, Innovation and Technology’s (<em>DSIT</em>) <em>AI Regulation White Paper</em> both stress a principle-led approach: safety, fairness, transparency, accountability, and contestability. Fine ideas — but without enforcement, they remain polite suggestions.</p><p>Still, the tide is turning. By late 2025, it’s expected that AI suppliers serving critical sectors (finance, health, infrastructure) will need to prove traceability — showing where every dataset, model, and dependency came from.</p><p>Supply-chain audits for AI are becoming the new “SOC 2” certification.</p><h3>The Cost of Ignorance</h3><p>According to IBM’s <em>Cost of a Data Breach 2025 Report</em>, incidents linked to third-party or supply-chain weaknesses now account for <strong>21 percent</strong> of all breaches globally. The average detection time? <strong>235 days.</strong> That’s eight months before an organisation realises a dependency failed them.</p><p>In the UK alone, data from the Information Commissioner’s Office (ICO) shows that <strong>17 percent of reported data incidents</strong> last year involved “unauthorised third-party access.” While not all were AI-related, the overlap is growing rapidly.</p><p>Why? Because AI tools are now embedded in everything — from HR portals to marketing dashboards. You don’t have to be running an “AI company” to have an AI supply chain. If your software uses automated summarisation, image analysis, translation, or sentiment detection — congratulations, you have one.</p><h3>The Transparency Mirage</h3><p>Many organisations think they’re safe because their vendor promises “no data sharing” or “enterprise-grade privacy.” But when you peel back the layers, you often find sub-processors — companies behind the companies.</p><p>A single AI integration might touch:</p><ul><li>Cloud infrastructure in Ireland,</li><li>An inference API in California,</li><li>An annotation service in India,</li><li>And an open-source dataset from who-knows-where.</li></ul><p>The paperwork might say “EU-only data storage,” yet logs might route through regions for load balancing. It’s not malice — it’s complexity. But complexity doesn’t look any different from negligence when regulators show up.</p><p>Transparency, the word everyone loves to print in brochures, often evaporates at the first audit.</p><h3>The Illusion of Control</h3><p>Here’s the uncomfortable truth: no one has full visibility.</p><p>Even the largest tech players rely on layers of external dependencies. When Microsoft launched Copilot, it wasn’t just Microsoft code at work — there were third-party language models, content filters, telemetry systems, and partner APIs.</p><p>Every layer adds capability but also uncertainty. Every patch fixes one issue but might introduce another.</p><p>Security teams try to map dependencies, but AI ecosystems shift too fast. A model version might change three times in a month, datasets update silently, APIs deprecate overnight. You can’t secure what you can’t pin down.</p><p>This is why the emerging mantra in cybersecurity circles is simple:</p><blockquote><em>“AI supply chain risk isn’t a flaw — it’s a fact.”</em></blockquote><p>The challenge isn’t to eliminate it, but to live with it intelligently.</p><h3>Numbers Don’t Lie</h3><p>Let’s bring in some context:</p><p>📊 <strong>Gartner (2025)</strong> — 78% of organisations adopting GenAI use at least one model or component from unverified third-party sources.</p><p>📊 <strong>NCSC</strong> — over 60% of UK companies deploying AI “lack full visibility of data provenance.”</p><p>📊 <strong>EY UK AI Trust Survey 2025</strong>–43% of executives admit they “do not fully understand their AI vendor’s architecture or data lineage.”</p><p>These numbers don’t reflect recklessness — they reflect speed. The corporate world ran toward AI faster than it built guardrails. Now we’re retrofitting safety into engines already in motion.</p><p>As one CISO told me at a recent event:</p><blockquote><em>“We were promised AI would automate our problems. Turns out, it just automated our risks.”</em></blockquote><h3>The Practical Playbook: How to Bring Shadows into the Light</h3><p>Alright, let’s get practical — because fear without action just leads to paranoia and more budget requests.</p><p>Here’s the playbook I’d recommend for every leader trying to stay sane in this messy AI supply chain world.</p><h4>1. Map Your AI Estate Like It Actually Exists, Not How It’s Documented</h4><p>Start by asking the uncomfortable questions:</p><ul><li>What models are we using?</li><li>Who built them?</li><li>Where do they live?</li><li>Who maintains them?</li></ul><p>Most companies can’t answer all four. Create a live registry — a <em>software bill of materials (SBOM)</em> for AI systems — listing every dataset, model, API, and plugin in play.</p><p>This one task, though boring, can reduce risk visibility gaps by over 40%, according to the UK’s NCSC.</p><p>Because you can’t secure what you can’t see.</p><h4>2. Verify Before You Trust</h4><p>Treat every AI component as “potentially compromised” until proven otherwise.</p><p>Run integrity checks, use hash verification, and source models only from reputable repositories.</p><p>Hugging Face now offers <em>model signing</em> to confirm authenticity — use it.</p><p>If a model vendor can’t explain their supply chain, that’s your red flag.</p><p>And yes, that includes cloud services that say “fully managed.”</p><p>Managed doesn’t mean monitored — it means <em>delegated</em>.</p><h4>3. Shift Left on Security</h4><p>The earlier you embed risk assessments, the cheaper your mistakes become. Before procurement approves a new AI vendor, make security part of the due diligence process.</p><p>NCSC’s <em>Secure AI Development Framework (SAIF)</em> has a neat recommendation: treat AI components like software dependencies — subject them to the same scanning, logging, and sandbox testing.</p><p>Or in simpler terms: test your models like you test your code.</p><h4>4. Contract Like You Mean It</h4><p>A good contract does more for AI security than three awareness sessions.</p><p>Include clear clauses around:</p><ul><li>Data residency and sub-processor disclosure.</li><li>Patch timelines for discovered vulnerabilities.</li><li>Rights to audit and demand model documentation.</li><li>Incident notification within 72 hours (yes, even for “beta” models).</li></ul><p>The ICO’s latest guidance hints that <em>“shared accountability does not dilute responsibility.”</em></p><p>In other words: if your vendor fails, it’s still your name in the headlines.</p><h4>5. Build Provenance Tracking</h4><p>Think of provenance as a “black box recorder” for your AI. Track every stage: data source → training → fine-tuning → deployment → update.</p><p>Tools like model cards, dataset datasheets, and lineage graphs can make this visual and auditable. You don’t need perfection; you need traceability.</p><p>Because when something goes wrong, it’s not just about fixing it — it’s about proving you knew how it went wrong.</p><h4>6. Create a Standing Red Team for AI</h4><p>Cyber teams red-team applications. Why not models? Simulate data poisoning, prompt injection, and dependency hijacks.</p><p>The <strong>AI Safety Institute</strong> in the UK has already started model red-teaming exercises with private companies. It’s a powerful shift from “compliance-driven audits” to “curiosity-driven assurance.”</p><p>Remember: you can’t out-regulate creativity — but you can prepare for it.</p><h4>7. Educate Your Humans</h4><p>Technology doesn’t fail in isolation — people do. Make sure your developers, data scientists, and procurement teams understand what AI supply-chain risk actually looks like.</p><p>Run short, scenario-based workshops: “What happens if our model is poisoned?” Get them thinking about risk, not just responsibility.</p><p>If you can’t change behaviour, you can’t secure culture.</p><h3>The UK Context: From Guidance to Governance</h3><p>The UK isn’t sitting idle — it’s just moving differently.</p><p>In 2025, the <strong>AI Safety Institute</strong> began collaborating with the <strong>National Cyber Security Centre (NCSC)</strong> and <strong>Department for Science, Innovation &amp; Technology (DSIT)</strong> to build a <em>“secure-by-design”</em> AI testing environment. This includes provenance verification, red-team testing, and certification templates for AI vendors.</p><p>The <strong>Digital Regulation Cooperation Forum (DRCF)</strong> — which unites the ICO, Ofcom, CMA, and FCA — is working on a shared approach for auditing AI supply chains.</p><p>The tone is pragmatic: less “thou shalt not,” more “here’s how to do it right.”</p><p>That’s very British — calm governance, gentle nudging, and just enough pressure to make you double-check your documentation.</p><p>But the reality is that UK organisations connected to EU or US markets will still have to align with stricter frameworks. Compliance boundaries don’t stop at Dover.</p><h3>Why This All Matters</h3><p>Because the future of AI won’t be defined by who builds the smartest model — but by who builds the <em>most trustworthy</em> one.</p><p>We’ve already learned that security and ethics can’t be afterthoughts. The next phase of AI maturity is about integrity — knowing where your systems come from and what they’re made of.</p><p>Think about it: you wouldn’t serve a meal without knowing the ingredients. Why would you serve an AI product without knowing its datasets?</p><p>In a world where AI drives financial decisions, medical recommendations, and customer interactions, provenance isn’t paperwork — it’s protection.</p><h3>My POV</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ABcRSdw84bUm6VKng9qI3Q.png" /></figure><p>I’ve been saying this for a while now — <strong>AI security isn’t really about defending algorithms; it’s about defending trust.</strong></p><p>We’ve built incredible systems that can reason, predict, and generate — but in the rush to innovate, we’ve started to treat trust as an optional feature rather than the foundation. Every model, every dataset, every plugin carries someone else’s fingerprints. Yet most organisations still assume “enterprise-grade” means “secure by default.” It doesn’t — it just means “expensive.”</p><p>What I see happening is simple: we’re outsourcing too much faith to tools we barely understand. We celebrate speed, but speed without visibility is just risk wearing sneakers.</p><p>If you strip away the hype, AI supply-chain risk comes down to one word — <strong>inheritance</strong>. When you adopt someone else’s model, you adopt their assumptions, their errors, and sometimes, their vulnerabilities.</p><p>So for me, the real differentiator won’t be who builds the most advanced models, but who builds the <em>most transparent ones</em>.</p><p>Because in the end, accountability is what will separate innovators from the rest.</p><h3>The Final Sip</h3><p>So, as we close this edition, here’s my takeaway for you to stir into your cup of tea: Every AI system has a family tree. Know yours.</p><p>Because the next big breach might not come from your code — it might come from someone else’s model that your team downloaded on a Friday.</p><p>The age of “move fast and break things” is over. The new era is “move wisely and verify everything.”</p><p>And trust me — that still counts as innovation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ijkBiwWhoGJ8D0j30NVCaQ.png" /></figure><p>☕️ <em>Thanks for reading “AI’s Shadow Supply Chain — The Hidden Risks Behind Every Prompt.”</em></p><p>If it helped shed some light on your own organisation’s shadows, share it with someone who still believes “our AI vendor handles that.”</p><p>Please don’t forget to leave a review.</p><p>If you found this helpful, <strong>subscribe</strong> to stay in the loop for future blogs.</p><p><strong>Together, let’s build a safer and more secure digital world! 💪🤗👨💻</strong></p><blockquote><strong>Get connected with me here :</strong></blockquote><blockquote><strong><em>LinkedIn: </em></strong><a href="https://www.linkedin.com/in/yash-gorasiya/"><em>https://www.linkedin.com/in/yash-gorasiya/</em></a></blockquote><blockquote><strong><em>Twitter: </em></strong><a href="https://twitter.com/r3v3Ncl4W"><em>https://twitter.com/r3v3Ncl4W</em></a></blockquote><blockquote><strong><em>Instagram: </em></strong><a href="https://instagram.com/r3v3ncl4w.go"><em>https://instagram.com/r3v3ncl4w.go</em></a></blockquote><blockquote><strong><em>Blog page: </em></strong><a href="https://instagram.com/forensis_digitum"><em>https://instagram.com/forensis_digitum</em></a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71fcfdf8038b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Compliance Hangover: When Innovation Outruns Regulation]]></title>
            <link>https://yashgorasiya.medium.com/the-ai-compliance-hangover-when-innovation-outruns-regulation-cca1e9ebf06a?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/cca1e9ebf06a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[regulation]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Tue, 21 Oct 2025 04:19:05 GMT</pubDate>
            <atom:updated>2025-10-21T04:19:05.163Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OTmnx5coIBKi1StpjQyLMw.png" /></figure><p>The party was glorious while it lasted. Executives toasted “transformational AI,” consultants promised infinite efficiency, and every press release suddenly had the word <em>Copilot</em> in it. Then the music stopped. Compliance woke up with a headache, Legal demanded receipts, and someone quietly asked the most inconvenient question in the room:</p><blockquote><em>“Did anyone check if we were actually allowed to do this?”</em></blockquote><p>Welcome to the AI compliance hangover — where yesterday’s innovation sprint meets today’s regulatory audit.</p><p>For the last two years, companies have treated AI like champagne at a corporate retreat: the bottle looked expensive, everyone wanted a glass, and no one asked who would pay for the next morning’s aspirin.</p><p>Now the regulators are at the door.</p><p>It’s good to have you back.</p><p>If you’ve been following our journey through the world of AI security and governance, you know we’ve talked about leaks, loopholes, and lessons.</p><p>This week, we’re tackling something every organisation is waking up to — the <strong>AI compliance hangover</strong>: that uncomfortable moment when the excitement of building fast meets the reality of being regulated even faster.</p><p>Before we get brewing, a quick note to everyone celebrating —</p><p>✨ <strong>Wishing you and your families a prosperous, bright, and safe Diwali!</strong> May your systems stay secure, your audits pass cleanly, and your firewalls sparkle brighter than your diyas.</p><p>Now, let’s pour into this week’s theme: when innovation runs ahead of the rulebook — and how to catch up before the regulators do.</p><h3>The Morning After</h3><p>Across boardrooms in London, New York, and Singapore, the mood has shifted from <em>build everything</em> to <em>explain everything.</em> Chief data officers who once bragged about deploying twenty-seven AI pilots are now quietly deleting half of them. Legal teams are discovering that “experimental” has no meaning in an audit, and HR is learning that “we didn’t know the model would say that” is not a valid defence under employment law.</p><p>The truth is simple: organisations moved faster than the rulebook. Regulators, never known for agility, are trying to catch up with a phenomenon that reinvented itself three times while they drafted the first paragraph.</p><p>In fairness, the hype made it easy to forget boring details like privacy notices, consent clauses, or cross-border data transfers. After all, when you’re promising the board 300 percent productivity gains, who wants to talk about paragraph 34 of the Data Protection Act?</p><h3>The Speed Problem</h3><p>AI innovation behaves like a Formula 1 car designed by a committee: engineered for speed, maintained by lawyers, and driven by whoever happens to shout “Go!” first.</p><p>By the end of 2024, more than 70 percent of UK enterprises had integrated some form of generative AI into their workflow, according to a PwC survey. Fewer than 30 percent had formal governance frameworks. The remaining 40 percent were apparently relying on “common sense” — a concept not yet defined in the ICO’s guidance notes.</p><p>This mismatch between speed and scrutiny created the modern compliance paradox: the faster you innovate, the more you expose. When the EU’s AI Act finally passed in 2024, many executives treated it like a foreign language exam they hadn’t studied for. Some still think “high-risk systems” refer to mountain-climbing equipment.</p><p>Meanwhile, the US issued an Executive Order on “Safe, Secure, and Trustworthy AI.” China introduced its Generative AI Measures. The OECD refreshed its AI Principles. Each promised harmony; together they created a symphony of conflicting obligations.</p><p>For global firms, this means one thing: every AI model is now a passport application.</p><h3>Innovation vs Regulation: A Global Tug-of-War</h3><p>If you map the world’s AI regulations, it looks less like a legal framework and more like abstract art.</p><p>The <strong>EU AI Act</strong> divides systems into risk categories: minimal, limited, high, and unacceptable. High-risk systems — think HR screening, credit scoring, or medical diagnostics — require impact assessments, transparency logs, and human oversight. Unacceptable systems, such as social-scoring engines, are banned outright.</p><p>In theory, this makes sense. In practice, it’s like running a triage ward with every patient claiming to be “medium risk.”</p><p>The <strong>United States</strong> takes a lighter approach: instead of one law, it offers a patchwork of sectoral guidance and an Executive Order that reads like a wish list — lots of ambition, limited enforcement.</p><p><strong>China’s</strong> rules are the opposite: prescriptive, fast-moving, and tightly controlled. Models must pass state security reviews before release, and anything that “undermines socialist values” is prohibited — a compliance clause with interesting philosophical implications.</p><p>And the <strong>United Kingdom?</strong> Ever the pragmatist, it chose a “pro-innovation” route. The DSIT White Paper proposed five flexible principles — safety, transparency, fairness, accountability, and contestability — and politely delegated implementation to existing regulators.</p><p>It’s the legislative equivalent of saying, “Everyone please behave,” and hoping for the best.</p><h3>The UK’s Balancing Act</h3><p>To its credit, the UK is trying to have its cake and regulate it too. The newly formed <strong>AI Safety Institute</strong> — tasked with testing frontier models for safety — has become the government’s answer to Silicon Valley’s speed. Meanwhile, the <strong>Information Commissioner’s Office (ICO)</strong> continues to remind companies that data-protection law already covers AI.</p><p>The <strong>National Cyber Security Centre (NCSC)</strong> publishes sober guidance on secure AI development, warning against prompt-injection attacks and data leakage. Most CISOs read it over coffee and nod gravely before returning to projects that violate at least three of its principles.</p><p>Then there’s the <strong>Digital Regulation Cooperation Forum (DRCF)</strong>, an alliance of UK regulators — the ICO, CMA, Ofcom, and FCA — trying to coordinate AI oversight. Imagine four referees explaining different rulebooks during the same match.</p><p>This is Britain’s approach in a nutshell: collaborative confusion with good intentions.</p><p>Still, the tone matters. While the EU AI Act threatens fines up to 7 percent of global turnover, the UK’s language is gentler, favouring guidance over punishment. Critics call it toothless; supporters call it proportionate. Most companies call it <em>unclear</em>.</p><h3>Corporate Reality: Compliance by PowerPoint</h3><p>Inside many organisations, AI governance has turned into a PowerPoint genre. There are slides titled “Ethical AI Vision,” followed by “Draft Policy (Version 0.3 — Pending Legal Review).” Someone presents them in a town-hall meeting, promising that “a comprehensive framework is coming soon.”</p><p>It rarely arrives.</p><p>One UK bank reportedly launched an AI ethics committee with twelve members and no budget. After six months, the committee produced a mission statement and a logo. Elsewhere, a global retailer created an “AI Risk Register” that lists 127 models but doesn’t specify who owns any of them.</p><p>This is the compliance hangover in action: documentation without discipline.</p><p>According to Deloitte’s 2025 Global AI Governance Survey, only 18 percent of organisations claim to have a “mature” governance structure for AI. Yet 81 percent plan to increase AI investment this year. In other words, we’re building faster cars and still arguing over road signs.</p><h3>The Confusion Economy</h3><p>We are, at present, living in what might be described as the “confusion economy” of AI. Boards are filled with optimism, compliance teams with anxiety, and regulators with open consultation surveys.</p><p>Ask five executives what “AI compliance” means, and you’ll get six answers. For some, it’s about ethics — “making sure the model doesn’t say anything offensive.” For others, it’s about legality — “ensuring GDPR boxes are ticked.” For most, it’s about survival — “doing enough to stay out of headlines.”</p><p>The problem is that AI isn’t one thing. It’s a stack of technologies — data, algorithms, models, APIs, user interfaces — all governed by overlapping jurisdictions. Trying to regulate “AI” as a single entity is like trying to regulate “electricity” by writing a policy about lightbulbs.</p><p>The result? A compliance industry racing to sell solutions for risks no one has yet defined. Workshops on “AI ethics,” courses on “AI assurance,” and an entire cottage economy of consultants offering to translate the EU AI Act into PowerPoint English.</p><h3>When the Rules Don’t Match the Reality</h3><p>Consider the paradox: regulators say they want <em>transparency</em>, but some of the world’s largest AI models are built on datasets so vast and complex that even their creators can’t fully explain how outputs are produced.</p><p>It’s a bit like requiring a Michelin chef to provide the chemical formula for “flavour.”</p><p>The EU AI Act demands that high-risk systems be “explainable and traceable.” Noble goal. Yet, when Google DeepMind released its Gemini Ultra model, its own engineers admitted that output reasoning remained “probabilistic and non-linear.”</p><p>The law says “explain your model.” The model says “that’s cute.”</p><p>This mismatch isn’t just academic. It creates real tension in corporate compliance. Lawyers ask data scientists for model documentation; data scientists respond with 800 pages of equations and a shrug. The compliance officer uploads the file to SharePoint and ticks “evidence retained.”</p><p>And so, the illusion of control continues.</p><h3>The Cost of Chaos</h3><p>According to IBM’s 2025 <em>Cost of Compliance Report</em>, AI-related regulatory fines are projected to surpass <strong>$1.3 billion globally</strong> by 2026. That’s not because regulators have suddenly become aggressive — it’s because most companies genuinely don’t know how to comply.</p><p>A PwC UK analysis found that 42 percent of British businesses implementing AI “have not conducted any form of risk or impact assessment.” Yet, under both the EU AI Act and UK GDPR, that’s a legal requirement for anything that processes personal data or influences human decisions.</p><p>When these gaps are discovered, the costs multiply: investigation fees, remediation plans, PR disasters, and lost trust. One insurance firm learned this the hard way when its AI claims triage system denied coverage based on biased historical data. The ICO didn’t fine them — but it did open an inquiry that froze the system for six months. The savings promised by automation evaporated in legal fees.</p><p>So yes, innovation moves fast — but compliance debt compounds faster.</p><h3>The “Innovation Defence”</h3><p>You’ll often hear executives justify grey-zone deployments with a familiar refrain: <em>“We can’t regulate what we don’t understand.”</em> This conveniently forgets that ignorance has never been a defence in law.</p><p>It’s the same logic teenagers use when caught speeding — “the sign wasn’t clear” — only this time the teenager is a multinational.</p><p>In fairness, the confusion isn’t always deliberate. Regulators themselves often contradict each other. In 2024, the European Data Protection Board said that AI profiling “should require explicit consent.” The EU AI Act, passed months later, referred to “appropriate safeguards” without defining them. Meanwhile, the ICO advised UK firms to perform “legitimate interest assessments.” Three regulators, one activity, three interpretations.</p><p>If compliance were a board game, no one would know whose turn it is.</p><h3>The New Arms Race: AI Assurance</h3><p>Out of this uncertainty emerges a new buzzword: <em>AI assurance.</em> It sounds impressive — somewhere between an insurance policy and a promise.</p><p>In practice, it’s the process of verifying that an AI system is safe, lawful, and aligned with declared principles. Think of it as the digital equivalent of a MOT test for algorithms.</p><p>The UK’s <strong>AI Standards Hub</strong>, backed by the British Standards Institution, has already started drafting frameworks. ISO is working on <strong>ISO/IEC 42001</strong>, an international AI management standard. The goal is to let companies say, “Our AI passed inspection.”</p><p>The irony, of course, is that the technology moves faster than the paperwork. By the time a certification is issued, the model may have been retrained, redeployed, or replaced.</p><p>It’s like issuing a safety certificate for a ship that’s already left the harbour.</p><p>Still, it’s progress. AI assurance may not eliminate risk, but it replaces blind optimism with structured accountability — and that’s a start.</p><h3>Case Studies: Learning the Hard Way</h3><h4>Case 1: The Overconfident Bank</h4><p>A global financial institution proudly deployed an AI tool to screen loan applications. The model, trained on ten years of internal data, replicated historic bias against certain postcodes. When flagged, the bank insisted its “AI ethics review” had cleared the system. The review, it turned out, was a single page titled “Model looks good.” The FCA stepped in, and the tool was quietly retired.</p><h4>Case 2: The Curious Retailer</h4><p>A retail chain used AI chatbots for customer service. The bot accidentally exposed order histories containing partial payment data. When asked about data handling, the company claimed “the chatbot doesn’t store information.” It did — on a third-party cloud. The ICO didn’t fine them but issued a public reprimand that cost far more in reputation.</p><h4>Case 3: The Ambitious Startup</h4><p>A London-based AI health startup trained its diagnostic model using anonymised NHS datasets. Unfortunately, its definition of “anonymised” didn’t match the NHS’s. Patient consent forms hadn’t covered commercial training use. Within months, the startup’s valuation halved as investors fled regulatory uncertainty.</p><p>Each of these stories ends the same way: ambition collided with ambiguity.</p><h3>The Human Cost of Compliance Fatigue</h3><p>Inside organisations, compliance fatigue is real. Teams juggling multiple frameworks — GDPR, ISO, SOC, AI Act — start to see governance as red tape, not risk management. Developers feel slowed down, business leads feel boxed in, and security officers feel like referees at a rugby match with no rules.</p><p>Yet ignoring compliance is like skipping medical check-ups. You feel fine until you don’t.</p><p>A survey by EY in 2025 found that <strong>61 percent of UK executives</strong> admitted their company “deploys AI faster than it can assess risk.” Of those, nearly half reported at least one “regulatory near-miss” in the past year.</p><p>The culture of “launch now, fix later” is becoming the new normal. Unfortunately, the “later” part often arrives sooner than expected — usually in the form of an ICO letter.</p><h3>Numbers Don’t Lie</h3><p>Statistics, as ever, are the sober friend at the end of the night.</p><p>According to Accenture’s <em>Global AI Readiness Index 2025</em>, only <strong>29 percent of organisations</strong> consider themselves “fully compliant” with upcoming AI regulations. Yet over <strong>84 percent</strong> already use AI in some form of decision-making.</p><p>IBM’s <em>Cost of Compliance 2025</em> estimates the average organisation now spends <strong>$5.2 million annually</strong> managing AI-related risk — a 38 percent increase from 2023. For every £1 spent on AI innovation, UK firms spend about 30p on compliance, a ratio that analysts say “feels optimistic.”</p><p>Meanwhile, McKinsey reports that <strong>two-thirds of executives</strong> believe regulation is “the primary barrier to scaling AI.” Translation: “We’d love to go faster if only someone else guaranteed it was safe.”</p><p>And that’s the paradox: companies want clear rules but resent them when they arrive. Everyone demands certainty, few welcome enforcement.</p><h3>The Practical Playbook (for the Morning After)</h3><p>So, what now? How do organisations nurse this collective hangover and get back to productive innovation — legally this time?</p><p>Think of compliance not as a checklist, but as a conversation between curiosity and caution.</p><p>Here’s the pragmatic version of that dialogue:</p><h4>1. Know what you’ve built</h4><p>Half the firms I speak with can’t list all the AI models running in their environment. Start there. Inventory before impact. You can’t govern what you can’t name.</p><h4>2. Perform the “good parent test”</h4><p>If your AI system made a questionable decision, could you explain to a regulator — or a journalist — how it reached that conclusion? If not, you need documentation, audit trails, and escalation processes.</p><h4>3. Adopt existing standards</h4><p>ISO/IEC 42001 (AI management systems), NIST’s AI Risk Framework, and the UK’s AI Standards Hub all exist to save you time. You don’t need to reinvent ethics; you just need to implement it.</p><h4>4. Build multidisciplinary governance</h4><p>Compliance isn’t a solo sport. Your AI oversight board should include legal, data, security, and business representation. The best models fail in testing, not in production.</p><h4>5. Train your humans</h4><p>No governance framework survives contact with an untrained employee. Make AI literacy part of induction. Explain what “responsible use” means. Fear doesn’t drive compliance — understanding does.</p><h4>6. Test like an adversary</h4><p>Run red-team exercises. Try to break your own AI. What happens when someone enters malicious prompts? Does it leak data, hallucinate, or generate bias? Learn your weaknesses before the press does.</p><h4>7. Document relentlessly</h4><p>If in doubt, write it down: why a model was built, what data trained it, how it’s monitored. Good documentation is boring — which is exactly why regulators love it.</p><h3>The UK Lens: Pragmatism with Purpose</h3><p>The UK continues to straddle two worlds: European compliance rigour and Anglo-American agility.</p><p>The Department for Science, Innovation and Technology (DSIT) plans to issue sector-specific guidance instead of one sweeping AI law. The <strong>AI Safety Institute</strong> has begun running “red team evaluations” of frontier models — a global first. And the <strong>ICO</strong> is working on templates for Data Protection Impact Assessments tailored to generative AI.</p><p>In other words, Britain’s approach is to measure twice and cut once. Slow, perhaps, but deliberate.</p><p>Contrast this with the EU’s rule-heavy framework or America’s laissez-faire pragmatism, and the UK looks oddly balanced — the responsible middle child of global AI governance.</p><p>Still, that doesn’t mean immunity. UK firms serving EU clients remain subject to the AI Act’s extraterritorial provisions. Those handling US data face cross-border transfer requirements. Compliance is now a passport, not a postcode.</p><h3>Why This Matters Beyond Checkboxes</h3><p>Because at its heart, this isn’t about rules; it’s about trust.</p><p>Trust that the data we share won’t be misused. Trust that decisions made by machines are fair and explainable. Trust that innovation doesn’t require recklessness.</p><p>If AI is to underpin the next industrial era, it needs a moral architecture as strong as its technical one. Without that, every breakthrough becomes another headline waiting to happen.</p><p>As an executive from a major insurer told me recently, “We’ve stopped asking what AI can do and started asking what we can defend in court.” That, right there, is the sound of the hangover.</p><h3>My POV</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*UXzxhiUK7n7rvwhi" /></figure><p>I’ve seen enough of these cycles to recognise the pattern. Every technological leap starts with euphoria — the sense that we’ve finally found the silver bullet. Then reality sets in.</p><p>AI is following the same script as cloud computing, social media, and fintech: move fast, break things, hire lawyers. The only difference is the scale of what’s at stake.</p><p>I don’t see compliance as a brake; I see it as a seatbelt. No one enjoys wearing it, but everyone’s grateful for it when the road bends unexpectedly.</p><p>If I were advising a board today, I’d say: Stop framing compliance as a cost centre and start treating it as competitive advantage. The companies that get this right early — the ones that can prove their AI is accountable — will be the ones customers, investors, and regulators trust to lead.</p><p>Because in 2025, credibility is currency.</p><h3>The Final Sip</h3><p>The AI revolution isn’t over; it’s just sobering up. The next phase won’t be about who built the fastest models, but who built the most trustworthy ones.</p><p>So, to all the leaders nursing their compliance hangovers this week: drink some water, review your risk register, and remember — the best innovation stories always end with accountability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*JzmAFJAICWS8W6Wc" /></figure><p>☕️ <em>Thanks for reading “The AI Compliance Hangover: When Innovation Outruns Regulation.”</em></p><p>If this helped clear your head (or your policy gaps), share it with someone still partying in the AI fast lane.</p><p>Please don’t forget to leave a review.</p><p>If you found this helpful, <strong>subscribe</strong> to stay in the loop for future editions.</p><p><strong>Together, let’s build a safer and more secure digital world! 💪🤗👨💻</strong></p><blockquote><strong><em>Get connected with me here :</em></strong></blockquote><blockquote><strong>LinkedIn: </strong><a href="https://www.linkedin.com/in/yash-gorasiya/">https://www.linkedin.com/in/yash-gorasiya/</a></blockquote><blockquote><strong>Twitter: </strong><a href="https://twitter.com/r3v3Ncl4W">https://twitter.com/r3v3Ncl4W</a></blockquote><blockquote><strong>Instagram: </strong><a href="https://instagram.com/r3v3ncl4w.go">https://instagram.com/r3v3ncl4w.go</a></blockquote><blockquote><strong>Blog page: </strong><a href="https://instagram.com/forensis_digitum">https://instagram.com/forensis_digitum</a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cca1e9ebf06a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Insider Threat 2.0 — When AI Makes Leaks Unintentional.]]></title>
            <link>https://yashgorasiya.medium.com/the-insider-threat-2-0-when-ai-makes-leaks-unintentional-b4d00e8c2515?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/b4d00e8c2515</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[infosec]]></category>
            <category><![CDATA[hacking]]></category>
            <category><![CDATA[insider-threat]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 06 Oct 2025 20:04:58 GMT</pubDate>
            <atom:updated>2025-10-06T20:04:58.198Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Insider Threat 2.0 — When AI Makes Leaks Unintentional.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v2onWufor3XMStO2OrlL1A.png" /></figure><p>You wouldn’t believe how it started. A tired engineer at the end of a twelve-hour shift copied a chunk of source code into ChatGPT. He only wanted help fixing a bug that had held his team up for days. Within hours, that same snippet was sitting in OpenAI’s training logs — potentially visible to hundreds of model trainers. A week later, Samsung issued a global memo: <em>All generative-AI use is banned until further notice.</em></p><p>That one small moment — human curiosity meeting limitless AI convenience — became a case study in how leaks no longer need malice. They just need a model eager to help.</p><p>Last week we explored EchoLeak, where AI itself was the attack vector. This week we’re looking inward. The threat isn’t a hacker — it’s us.</p><p>For decades, the phrase <em>insider threat</em> meant a disgruntled employee, an espionage recruit, or someone selling data on the dark web. In 2025, the definition has changed. Now, an insider can simply be an overworked marketer using Copilot to polish a proposal, or a lawyer using Gemini to draft a contract — unaware that confidential clauses are being logged, cached, or learned from.</p><p>The enemy isn’t intent; it’s ignorance paired with automation.</p><h3>How Accidental Leaks Happen</h3><p>Generative AI tools are designed to absorb context. They work by consuming whatever you feed them — your text, your questions, your tone — and turning it into patterns that help them predict the next word. The problem: those patterns sometimes include sensitive data.</p><p>There are four main ways leaks occur:</p><ol><li><strong>Shadow AI</strong> — employees use public AI tools outside approved systems because they’re convenient. It’s the digital version of “I’ll just send it to my Gmail.” According to Gartner’s 2024 workplace survey, 59 % of employees admitted using unapproved AI tools at work.</li><li><strong>Context creep</strong> — tools like Copilot and Gemini search across internal drives to “personalise” responses. That context can include draft contracts, client data, or even HR records.</li><li><strong>Fine-tuning leaks</strong> — if a company uses internal data to train or improve a model without proper isolation, that data can bleed into outputs for other users.</li><li><strong>Prompt memory</strong> — chat histories stored on cloud servers may contain sensitive text, and even if deleted later, logs often persist for compliance or model improvement.</li></ol><p>It’s like hiring an assistant with photographic memory who never forgets anything you tell them — and then lending that assistant to the rest of the world.</p><p>Generative AI tools are designed to absorb context. They work by consuming whatever you feed them — your text, your questions, your tone — and turning it into patterns that help them predict the next word. The problem: those patterns sometimes include sensitive data.</p><p>There are four main ways leaks occur:</p><ol><li><strong>Shadow AI</strong> — employees use public AI tools outside approved systems because they’re convenient. It’s the digital version of “I’ll just send it to my Gmail.” According to Gartner’s 2024 workplace survey, 59 % of employees admitted using unapproved AI tools at work.</li><li><strong>Context creep</strong> — tools like Copilot and Gemini search across internal drives to “personalise” responses. That context can include draft contracts, client data, or even HR records.</li><li><strong>Fine-tuning leaks</strong> — if a company uses internal data to train or improve a model without proper isolation, that data can bleed into outputs for other users.</li><li><strong>Prompt memory</strong> — chat histories stored on cloud servers may contain sensitive text, and even if deleted later, logs often persist for compliance or model improvement.</li></ol><p>It’s like hiring an assistant with photographic memory who never forgets anything you tell them — and then lending that assistant to the rest of the world.</p><h3>Real Stories That Should Make You Pause</h3><p>The Samsung case was only the headline.</p><p>At Amazon, employees were warned not to paste confidential code into ChatGPT after snippets resembling proprietary algorithms appeared in model outputs.</p><p>At a London-based law firm, a paralegal reportedly used a public LLM to “summarise” witness statements; portions of those summaries later surfaced in unrelated AI-generated drafts.</p><p>And a university in the Midlands discovered students’ essays, uploaded for plagiarism checks through an external AI tool, were being indexed on public web caches.</p><p>None of these people were malicious.</p><p>They were simply trying to save time.</p><p>But each small act created an uncontrolled copy of private data.</p><h3>Why It Matters More Than You Think</h3><p>The immediate risk is obvious: confidential information leaves your environment.</p><p>But the long-term impact is bigger — reputation, regulation, and trust.</p><p>From a business perspective, insider leaks now represent nearly <strong>25 % of all data-loss incidents</strong> (IBM 2024). From a financial one, insider-related breaches cost companies <strong>over $11 million on average</strong>, almost twice the cost of external attacks.</p><p>And from a legal angle, the UK’s <em>Data Protection Act 2018</em> and GDPR equivalents make organisations — not vendors — responsible for what happens to personal data.</p><p>That means “the model did it” is not a valid defence.</p><p>If an employee feeds personal or client data into a public LLM, regulators still hold the employer accountable.</p><p>Then comes the human layer.</p><p>Employees start to fear AI surveillance; customers start to distrust brands that can’t safeguard data. In YouGov’s 2024 survey, only <strong>28 % of UK adults</strong> said they trusted companies to use AI responsibly. Every accidental leak pushes that number lower.</p><h3>The Technical Anatomy of an Unintentional Leak</h3><p>Let’s strip the jargon.</p><p>An LLM — large language model — learns from text. When you prompt it, the text goes into a temporary memory called a <em>context window</em>.</p><p>Depending on the product, that data might:</p><ul><li>stay local (in enterprise setups),</li><li>be cached for a few hours, or</li><li>be stored for model improvement.</li></ul><p>If those logs are later used for retraining, fragments of the original text can reappear in outputs — a phenomenon researchers call <em>data regurgitation.</em></p><p>That’s how personal names or code snippets sometimes leak from one user’s query to another’s.</p><p>Even enterprise versions promise privacy, but “no data retention” often means “limited retention for debugging.” It’s the grey area between marketing and reality.</p><p>Imagine whispering a secret to someone who promises to forget it right away — but has a habit of sleep-talking.</p><h3>The Scale of the Problem</h3><p>Let’s talk numbers again.</p><p>In 2025, roughly <strong>71 % of UK organisations</strong> reported experimenting with generative AI internally, but fewer than <strong>30 %</strong> had formal governance policies (PwC Digital Trust 2025).</p><p>Globally, <strong>one in five</strong> firms experienced an accidental data exposure linked to AI tools within the past year (IBM X-Force 2025).</p><p>And it isn’t slowing down.</p><p>Spending on AI productivity tools is up 230 % year-on-year across UK SMEs, while investment in AI security and governance has only risen 40 %.</p><p>We’re buying faster cars and still arguing about seatbelts.</p><h3>The UK’s Tightrope Walk</h3><p>The Information Commissioner’s Office (ICO) has issued repeated guidance reminding companies that AI use falls squarely under existing data-protection law.</p><p>If an employee uploads personal data to a generative-AI tool, that’s a data transfer to a third party — full stop.</p><p>The National Cyber Security Centre (NCSC) echoed this in early 2025, urging public-sector bodies to “treat generative-AI tools as untrusted external systems unless formally assessed.” Several councils have since banned staff from using open-AI interfaces for citizen data.</p><p>Private sector firms are responding unevenly.</p><p>Banks like Barclays and HSBC introduced internal “AI sandboxes” where employees can safely use models trained on synthetic data.</p><p>The NHS is piloting similar systems for clinical summarisation, but progress is slow due to privacy concerns.</p><p>For UK organisations, the tension is constant: adopt fast and risk leaks, or move slowly and lose the efficiency race.</p><h3>The Human Element That Never Went Away</h3><p>Here’s the irony: AI was supposed to solve the human problem — cut out error, fatigue, distraction. Yet here we are, circling back to the same root cause. It’s still us.</p><p>What’s changed is <em>how</em> our mistakes manifest. Ten years ago, a careless employee might’ve sent an email to the wrong client. Today, that same employee might ask an AI assistant to “summarise” client proposal X” — unknowingly including confidential financial details in a prompt that leaves the organisation’s network.</p><p>The psychology behind it is fascinating. Generative AI feels <strong>private</strong>. You’re typing to what looks like a blank screen — not another person. The barrier of embarrassment disappears. We over-share. We trust too easily.</p><p>Researchers at Oxford’s Internet Institute found that people are <strong>twice as likely</strong> to share sensitive data with conversational interfaces than via traditional software forms. Why? Because a chatbot “feels” like it’s on your side. That empathy illusion is powerful — and risky.</p><p>So the insider threat hasn’t disappeared; it’s just wearing a friendlier face.</p><h3>When “Helpful” Becomes Harmful</h3><p>There’s another subtlety few talk about: not all AI leaks are explicit. Some are <strong>derived</strong>.</p><p>Picture this. You’re in a product meeting. You ask Copilot, “Generate a summary of our latest R&amp;D direction.” Copilot pulls context from SharePoint documents, meeting notes, and internal files. You share the summary with a vendor — clean and well-written.</p><p>But hidden within that summary are references and phrasings unique to an unreleased prototype. You’ve just leaked strategy — not code, not data, but <strong>intent</strong>.</p><p>That’s what makes this new generation of insider threats so complex. It’s not just about sensitive data being uploaded — it’s about <strong>sensitive meaning</strong> being rephrased and redistributed by AI.</p><p>And once it’s out, good luck proving how it left.</p><h3>Lessons from the Past, Warning for the Future</h3><p>If EchoLeak was about AI being exploited, Insider Threat 2.0 is about <strong>us exploiting AI without understanding it</strong>.</p><p>The most common refrain I hear from CISOs lately is, “We can’t secure what we can’t see.” Shadow AI is exactly that — invisible technology usage happening inside visible organisations.</p><p>In IBM’s 2025 <em>Insider Risk Report</em>, 43% of UK enterprises admitted they don’t have a clear inventory of which departments use which AI tools. That’s like running a factory without knowing which machines are connected to power.</p><p>The future of data protection won’t just be about firewalls or encryption — it’ll be about <strong>visibility</strong>. Who used what, for what purpose, and with what data.</p><p>The companies that figure that out first will be the ones regulators trust most — and customers too.</p><h3>The Numbers Behind the Curtain</h3><p>Let’s make it concrete.</p><p>According to Verizon’s <em>2024 Data Breach Investigations Report</em>, internal actors were involved in <strong>19% of breaches</strong> globally. Of those, nearly half were <strong>accidental</strong> — not malicious.</p><p>The Ponemon Institute’s 2025 study on insider threats estimated the <strong>average annual cost per organisation</strong> at <strong>$11.5 million</strong>, up 40% since 2020. The same report found that <strong>AI-related data exposure</strong> accounted for nearly a fifth of new insider incidents in 2025 — a category that didn’t even exist three years ago.</p><p>On UK soil, the ICO reported a 17% increase in “data disclosed in error” cases in the last year alone, much of it linked to automated tools.</p><p>Numbers tell a story: we’re creating more data, training more models, and building more leaks — unintentionally.</p><h3>What Organisations Can Actually Do</h3><p>Let’s be honest: you can’t stop people from using AI. The convenience is too great, the productivity gains too real. But you <em>can</em> build guardrails so the benefits don’t become liabilities.</p><p>Here’s what I’ve seen work across different organisations I’ve interacted with:</p><h4>1. Acknowledge reality</h4><p>Stop pretending employees aren’t using public AI tools. They are. Conduct anonymous surveys, find out which ones are most popular, and meet the need officially instead of ignoring it.</p><h4>2. Create safe AI zones</h4><p>Establish internal LLMs or secure enterprise integrations where prompts and data stay within your control. Barclays’ internal AI sandbox is a good example.</p><h4>3. Educate differently</h4><p>Traditional security training doesn’t cut it anymore. “Don’t click suspicious links” needs to become “Don’t share data with systems you don’t own.” Make it relatable — run exercises showing how easy it is for context to leak.</p><h4>4. Policy, not paranoia</h4><p>A clear AI usage policy shouldn’t read like a list of bans. It should guide safe behaviour: what can be shared, what’s confidential, which systems are approved. Make it empowering, not intimidating.</p><h4>5. Build visibility</h4><p>Use AI governance tools that log prompts, data sources, and access. It’s not about spying — it’s about traceability when something goes wrong.</p><h4>6. Integrate with governance</h4><p>Your AI policies should tie directly into data governance, risk management, and compliance frameworks. That way, when the ICO comes calling, you’re not scrambling for documentation.</p><h3>The UK’s Moment of Reckoning</h3><p>In the UK, we’ve always had a pragmatic relationship with technology. We adopt quickly but regulate cautiously. EchoLeak pushed us to rethink AI security. The rise of insider leaks is forcing a rethink of <strong>AI privacy</strong>.</p><p>The ICO’s 2025 position paper makes one thing clear: “Responsibility for lawful AI use rests with the organisation deploying it.” That includes employee misuse. The NCSC has gone further, releasing practical playbooks for “Secure AI Implementation in the Public Sector.”</p><p>Meanwhile, Parliament’s <em>All-Party Parliamentary Group on AI</em> has been debating whether to classify generative AI leaks as reportable incidents under the UK GDPR framework — the same as ransomware.</p><p>If that happens, AI leak disclosures could become public record. Imagine explaining to the press that your company’s “AI assistant accidentally shared client data.” It’s not just a compliance issue; it’s a reputational one.</p><h3>My POV</h3><p>I’ve said this before: AI doesn’t have to be malicious to be dangerous.</p><p>When I talk to leaders, I can feel the tension. They want innovation but fear losing control. Here’s the uncomfortable truth — control was never the goal. <strong>Understanding</strong> should be.</p><p>AI is like a mirror. It reflects our habits, our shortcuts, our blind spots. When people paste confidential text into an AI chat, they’re not being careless — they’re being human. They’re seeking help. Our job as security and governance professionals is to make sure help doesn’t come with hidden costs.</p><p>If I could leave you with one idea, it would be this: <em>the new insider threat isn’t a person, it’s a pattern.</em> It’s the pattern of convenience over caution, of tools over principles. And fixing it doesn’t start with fear. It starts with culture — one where curiosity is welcome, but boundaries are clear.</p><p>So next time someone says, “I just asked AI for a quick suggestion,” pause and ask: <em>what did it learn from us in return?</em></p><h3>The Final Sip</h3><p>The insider threat used to be a shadow in the corner of every organisation. Now, it’s the glow on every employee’s screen.</p><p>AI has changed the game — and so must we. The goal isn’t to ban, but to balance; not to limit, but to lead. Because at the end of the day, AI will only be as responsible as the people who use it.</p><p>☕️ <em>Thanks for reading this edition of AI x Securi-TEA. If it got you thinking, share it with someone who still believes “the real risks are all external.”</em></p><p>Please don’t forget to leave a review.</p><p>And hey, if you found this helpful, <strong>subscribe</strong> to stay in the loop for future blogs.</p><p><strong>Together, let’s build a safer and more secure digital world! 💪🤗👨💻</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b4d00e8c2515" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[EchoLeak: When Your AI Copilot Becomes the Hacker’s Backdoor]]></title>
            <link>https://yashgorasiya.medium.com/echoleak-when-your-ai-copilot-becomes-the-hackers-backdoor-a5497816da50?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5497816da50</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[microsoft-copilot]]></category>
            <category><![CDATA[hacking]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 29 Sep 2025 10:05:00 GMT</pubDate>
            <atom:updated>2025-09-29T10:05:00.774Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Trt2hSLjo-u685_NlndrDQ.png" /></figure><p>Hey everyone 👋, welcome back to another edition of AI x Securi-TEA.</p><p>First, a quick cheers — this is our <strong>11th cup together</strong> ☕️. Over the past editions, we’ve gone from exploring why humans still matter in AI security, to how governance acts as the brakes that let us innovate safely, and why businesses can’t afford to treat AI as a shiny toy without accountability.</p><p>This week, though, we’re turning to something that gave me pause. Something that isn’t a “what if” scenario or a theoretical risk. It actually happened.</p><p>It’s called <strong>EchoLeak</strong> — a <strong>zero-click vulnerability in Microsoft 365 Copilot</strong> discovered in 2025.</p><p>What’s so shocking? You didn’t need to click a dodgy link. You didn’t need to download a shady attachment. You didn’t even need to make a mistake.</p><p>All an attacker had to do was send an email. And suddenly, your AI assistant — the one wired into your Outlook, Teams, OneDrive, and company data — could be tricked into handing over secrets.</p><p>No phishing. No malware. Just trust misplaced.</p><p>So today, let’s unpack what happened, why it matters, what organisations can learn from it, and how I personally think we need to adapt if we don’t want to see more “EchoLeaks” in tomorrow’s headlines.</p><h3>The EchoLeak Incident: What Really Happened</h3><p>In mid-2025, researchers disclosed <strong>EchoLeak (CVE-2025–32711)</strong> — a vulnerability in Microsoft 365 Copilot.</p><p>For those who haven’t used it, Copilot is Microsoft’s big bet: an AI assistant built directly into Word, Outlook, Excel, PowerPoint, and Teams. Instead of juggling tabs, you just “ask Copilot.” Draft me an email. Summarise this meeting. Analyse these sales trends. It’s marketed as the AI helper that never sleeps.</p><p>And it really is powerful. But EchoLeak showed us the shadow side.</p><p>Here’s how it worked:</p><ol><li><strong>Crafted an email with hidden instructions.</strong> Attackers embedded malicious instructions inside emails using Markdown links, HTML comments, or even image tags. These weren’t visible to normal users.</li><li><strong>Copilot preloads context.</strong> When asked to help with tasks, Copilot pulled in the email text (and those hidden instructions) as part of its “context.”</li><li><strong>Bypassing Microsoft’s defences.</strong> Copilot has systems like <strong>XPIA (Cross Prompt Injection Attempt classifiers)</strong> to detect malicious prompts. EchoLeak sidestepped them. It also dodged link-rewriting systems and content security checks.</li><li><strong>Privilege escalation through proxies.</strong> Attackers exploited Teams’ proxy behaviour and auto-fetched resources, using Microsoft’s own infrastructure as a trusted relay.</li><li><strong>Exfiltration.</strong> The AI, tricked by hidden commands, pulled sensitive internal data (emails, files, calendars) and sent it outside the company — to attacker-controlled servers.</li></ol><p>📌 All this happened without the victim clicking a thing.</p><p>To translate for non-technical readers: imagine hiring a personal assistant. Someone sends them a letter with invisible ink that says, “Make a copy of your boss’s diary and leave it under the doormat.” The assistant, being too eager, follows those hidden instructions without asking. That’s EchoLeak.</p><h3>Why EchoLeak Is Such a Big Deal</h3><p>Normally, when we talk about cyberattacks, the weak link is us — the humans.</p><ul><li>Phishing emails work because someone clicks.</li><li>Ransomware spreads because someone opens an attachment.</li><li>Business email compromise (BEC) works because someone trusts the wrong invoice.</li></ul><p>That’s why so much effort goes into security awareness training.</p><p>But EchoLeak flipped the script:</p><ul><li><strong>No human mistake needed.</strong> The AI was the weak link, not the employee.</li><li><strong>Everyday tools became attack vectors.</strong> Microsoft 365 is used by millions of UK businesses — from law firms to NHS trusts. A vulnerability here isn’t niche, it’s catastrophic.</li><li><strong>Blast radius is massive.</strong> Copilot isn’t just reading your email. It’s tied to Teams chats, SharePoint docs, OneDrive files, even internal business apps. Once compromised, the scope of what leaks grows fast.</li></ul><p>And here’s where numbers speak louder:</p><ul><li>According to Concentric AI, Microsoft Copilot already has access to an average of <strong>3 million sensitive records per organisation</strong>. Think HR files, financial reports, contracts. (<a href="https://www.techradar.com/pro/microsoft-copilot-has-access-to-three-million-sensitive-data-records-per-organization-wide-ranging-ai-survey-finds-heres-why-it-matters?utm_source=chatgpt.com">TechRadar Pro, 2025</a>)</li><li>IBM’s <em>Cost of a Data Breach 2023</em> report pegged the average cost of a breach at <strong>$4.45M</strong>. AI-related breaches? They’re even more expensive due to complexity.</li><li>In the UK, <strong>32% of businesses reported a cyber breach in 2023</strong> (UK Gov Cyber Security Breaches Survey). EchoLeak-style flaws will only push that number higher.</li></ul><p>📌 Translation: EchoLeak isn’t “just another bug.” It’s a preview of the next decade of cyber risk.</p><h3>What Went Wrong</h3><p>The beauty and danger of EchoLeak lies in the fact that it wasn’t one giant mistake. It was a chain of small gaps, each of which attackers exploited. Think of it like leaving a series of windows half-open in a skyscraper: one by itself isn’t catastrophic, but together, they form a perfect entry path.</p><p>Here’s the breakdown:</p><h3>Weak Prompt Filtering</h3><ul><li>Copilot had systems (like Microsoft’s <strong>XPIA classifiers</strong>) designed to detect malicious instructions.</li><li>But attackers used sneaky tricks — like <strong>Markdown links</strong> or <strong>HTML comments</strong> — to hide harmful prompts.</li><li>The filters looked for obvious “bad words” but missed the disguised ones.</li></ul><p>📌 <em>Analogy:</em> It’s like airport security that checks bags for “guns” but misses a weapon labelled “metal tool.”</p><h3>Over-Trusting External Content</h3><ul><li>Copilot automatically fetched images, links, and other resources referenced in emails and documents.</li><li>The assumption: “If it’s in your email, it must be safe.”</li><li>But attackers abused this by hiding malicious payloads in seemingly harmless content.</li></ul><p>📌 <em>Analogy:</em> Like your assistant accepting every free sample handed to them on the street — and bringing it straight into the office.</p><h3>Proxy Behaviour Exploited</h3><ul><li>Teams and Outlook use Microsoft proxies (servers that relay content).</li><li>Some of these proxies were too trusting — they didn’t fully validate where requests originated.</li><li>Attackers exploited this to escalate privileges and make Copilot treat their content as “trusted.”</li></ul><p>📌 <em>Analogy:</em> A delivery driver flashes a fake company badge, and security waves them through because they “look official.”</p><h3>No Strict Segmentation</h3><ul><li>Once malicious instructions entered one part of the system, they weren’t isolated.</li><li>The AI pipeline treated all instructions (system, user, external) as part of the same “conversation.”</li><li>This allowed attackers to hop between trust zones — from reading an email to leaking OneDrive files.</li></ul><p>📌 <em>Analogy:</em> Like a hotel with no locks between guest floors. Once you’re inside, you can wander everywhere.</p><h3>Lack of Adversarial Testing</h3><ul><li>Microsoft had strong QA for software bugs. But prompt injection? Not enough.</li><li>No one had thoroughly tested “what happens if someone hides instructions in clever ways.”</li><li>Attackers simply tested harder than defenders.</li></ul><p>📌 <em>Analogy:</em> You reinforced the front door, but never asked, “What if someone climbs the fire escape?”</p><p><strong>The takeaway:</strong> EchoLeak wasn’t about genius hackers. It was about small cracks — weak filters, too much trust, missing isolation — adding up. Attackers just followed the cracks until the whole wall gave way.</p><h3>Real-World Parallels</h3><p>EchoLeak may feel new, but it echoes (pun intended) past lessons:</p><ul><li><strong>SolarWinds (2020):</strong> Attackers slipped malicious code into trusted updates, breaching 18,000 orgs. Trust without verification = disaster.</li><li><strong>MOVEit Hack (2023):</strong> One vulnerable vendor led to thousands of victims — from BBC to BA. A reminder that a single weak spot can ripple globally.</li><li><strong>DeepMind &amp; NHS (2017):</strong> Not an attack, but a governance fail. Millions of patient records used without consent. Showed the dangers of moving fast without rules.</li><li><strong>Deepfake CEO scam (UK, 2019):</strong> Criminals cloned a CEO’s voice to trick a manager into wiring £200k. Proved how AI can supercharge old scams.</li></ul><p>📌 Lesson across all: whether it’s software updates, supply chains, AI assistants, or voices — attackers exploit <strong>trust</strong>. And we’re still too quick to give it.</p><h3>Numbers Don’t Lie</h3><ul><li><strong>74% of breaches involve the human element</strong> (Verizon DBIR 2023). For years, this stat framed phishing as the biggest threat. EchoLeak flips that on its head. It tells us: <em>what if the human isn’t even in the loop anymore?</em> We can’t train our way out of AI-native attacks.</li><li><strong>97% of AI breaches lacked proper access controls</strong> (IBM 2025). That means in almost every case, attackers got in not because they were brilliant, but because basic security wasn’t applied. EchoLeak is a textbook example — Copilot had too much access, too much trust, and too few boundaries.</li><li><strong>20% of organisations feel “highly prepared” for AI risks</strong> (PwC 2024). Think about that. If you walked into a boardroom of 10 companies, only two would say, “We’re confident in handling AI risks.” The other eight are effectively flying blind.</li><li><strong>28% of UK adults trust companies to use AI responsibly</strong> (YouGov 2024). Trust is already fragile. Incidents like EchoLeak make people ask: “If they can’t even stop an AI assistant from being tricked, why should I trust them with my data?”</li></ul><p>📌 The big picture: the numbers aren’t abstract. They show us a dangerous gap — AI adoption is skyrocketing, but AI risk management is crawling. That gap is where attackers thrive.</p><h3>Practical Playbook</h3><p>Let’s get practical. If you’re running AI tools like Copilot in your organisation, here’s a <strong>layered roadmap</strong> you can actually act on:</p><h3>Immediate Actions (This Week)</h3><ul><li><strong>Turn off risky defaults.</strong> Disable auto-fetching of external content (images, links) unless you really need it. Most orgs don’t even know this is enabled.</li><li><strong>Map Copilot permissions.</strong> Run a quick audit: what files, chats, and systems does Copilot touch? You might be shocked at how much it can see.</li><li><strong>Strip suspicious inputs.</strong> Add a sanitisation layer before prompts hit your AI. Remove hidden markdown, comments, or embedded links.</li></ul><p>📌 Why it matters: Most EchoLeak-style attacks don’t need sophistication — they exploit things left “on by default.” Quick checks reduce the low-hanging fruit.</p><h3>Short-Term (This Month)</h3><ul><li><strong>Red-team your AI.</strong> Just like you pen-test apps, run adversarial prompts against your AI. Can you trick it into revealing things it shouldn’t?</li><li><strong>Trust boundary mapping.</strong> Document every place where AI moves between systems. Where does Outlook talk to SharePoint? Where does Teams connect to OneDrive? Every “hop” is a risk.</li><li><strong>Provenance-based controls.</strong> Only allow AI to act on content when the source is verified. If you don’t know where a link came from, don’t trust it.</li></ul><p>📌 Why it matters: EchoLeak chained multiple weak assumptions. Red-teaming and trust mapping help you find those chains before attackers do.</p><h3>Strategic (This Quarter and Beyond)</h3><ul><li><strong>AI risk dashboards.</strong> Track metrics like: how many adversarial prompts fooled your AI? How fast did you patch? How many permissions are over-broad? Treat these like uptime SLAs.</li><li><strong>Bake into governance.</strong> Make AI risk part of your compliance framework. Not “extra,” but a core part of audits and board reporting.</li><li><strong>Human kill-switch.</strong> Build protocols where humans can pause or override AI activity if suspicious behaviour is detected. Think of it as the circuit breaker in your AI system.</li></ul><p>📌 Why it matters: AI risk isn’t going away. Building governance + security into the DNA of your organisation ensures you’re not reacting in panic mode later.</p><h3>The UK Context</h3><p>EchoLeak isn’t just a Microsoft problem. It’s a <strong>UK business problem</strong> because of how deeply tied we are to Microsoft’s ecosystem.</p><ol><li><strong>Microsoft 365 dominance.</strong> Over <strong>90% of UK enterprises</strong> use Microsoft 365 for collaboration (Statista 2024). That means if Copilot is compromised, the blast radius covers schools, councils, law firms, NHS trusts, and banks. This isn’t fringe tech — it’s national infrastructure.</li><li><strong>ICO Guidance (2023).</strong> The Information Commissioner’s Office set out that AI must be fair, transparent, accountable, and explainable. Imagine standing before the ICO and explaining EchoLeak: <em>“Yes, our AI leaked patient records, but it wasn’t our fault — it was tricked by hidden markdown.”</em> That wouldn’t fly.</li><li><strong>NCSC + CISA Secure AI Guidelines (2023).</strong> The National Cyber Security Centre, alongside US CISA, already warned that AI pipelines are new attack surfaces. They even listed prompt injection and model manipulation as high-risk threats. EchoLeak is literally the case study they were preparing us for.</li><li><strong>EU AI Act ripple effect (2024).</strong> Even though the UK isn’t bound by EU law, any firm serving European clients must comply. The AI Act is strict: high-risk systems (finance, healthcare, policing) need human oversight, audits, and explainability. EchoLeak shows why — because “black box AI” can’t be left unchecked.</li><li><strong>Public trust gap.</strong> Only <strong>28% of UK adults trust AI use by companies</strong> (YouGov 2024). Trust is a competitive advantage. Lose it, and customers won’t share data, employees won’t adopt tools, and regulators will clamp down harder.</li></ol><p>📌 The UK angle is clear: EchoLeak is a test case for how regulators, businesses, and the public see AI. Fail to learn from it, and we’ll see heavier-handed rules — and higher fines.</p><h3>My POV</h3><p>When I first read about EchoLeak, my immediate thought was: <em>this is our SolarWinds moment for AI.</em> Not because the technical details are the same, but because it exposed something we all assumed was safe and turned that assumption upside down.</p><p>We’ve been treating AI assistants like “smart interns”, giving them access, letting them help, and trusting they’ll behave. EchoLeak reminded me that these interns can be manipulated, and when they are, it’s not just embarrassing, it’s dangerous.</p><p>Here’s how I see it:</p><ol><li><strong>AI is not “just another feature.”</strong> Too many organisations still think of AI as an add-on, like a plug-in for Outlook or a shiny new analytics dashboard. However, EchoLeak reveals that AI is a <strong>new attack surface altogether</strong>. If you’re not building specific defences for it, you’re already behind.</li><li><strong>You can’t outsource accountability.</strong> If your Copilot leaks data, your customers won’t say “oh, Microsoft messed up.” They’ll say <em>you</em> messed up. Regulators will knock on <em>your</em> door. Accountability doesn’t travel with the vendor contract — it sits with the organisation that deployed the AI.</li><li><strong>Governance must evolve.</strong> For years, governance was about data privacy, fairness, and explainability. That’s still important — but now, governance must also include <strong>prompt testing, provenance checks, and AI threat modelling.</strong> Without that, you’ve got policy on paper and exposure in practice.</li><li><strong>Humans are still critical.</strong> We’ve been told AI will replace boring tasks and take humans “out of the loop.” EchoLeak shows why that’s a fantasy. You need humans who can step in, hit the kill switch, and ask the uncomfortable questions: <em>“Why is our AI pulling this data?”</em> AI can scale, but only humans can decide when to stop.</li></ol><p>If I were advising a board tomorrow, here’s the question I’d ask them directly: 👉 <em>“What could our AI be tricked into doing — and what are we doing today to stop that?”</em></p><p>Because here’s my belief: every organisation will face an EchoLeak-style moment. The difference between headlines and footnotes will be whether you prepared for it.</p><p>📌 My line: <strong>AI expands the attack surface. Governance secures it. But humans must always hold the final gate.</strong></p><h3>Final Sip</h3><p>EchoLeak is a reminder that AI security and AI governance are two sides of the same coin.</p><p>The winners of the AI era won’t just deploy AI assistants faster. They’ll be the ones who govern them tightly, secure them thoroughly, and keep humans firmly in control.</p><p>Innovation without defence isn’t progress. It’s just exposure on autopilot.</p><h3>❤️ Thanks for Reading</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/0*O3jQEuhUGnI3Mj9x.png" /></figure><p>Thanks for sticking with me till here! 🙏</p><p>Please don’t forget to leave a review or share your thoughts in the comments <strong>🥳🙏🫶 </strong>.</p><p>And hey, if you found this helpful, <strong>subscribe</strong> to stay in the loop for future blogs.</p><p><strong>Together, let’s build a safer and more secure digital world! 💪🤗👨💻</strong></p><blockquote><strong>Get connected with me here :</strong></blockquote><blockquote><strong><em>LinkedIn: </em></strong><a href="https://www.linkedin.com/in/yash-gorasiya/"><em>https://www.linkedin.com/in/yash-gorasiya/</em></a></blockquote><blockquote><strong><em>Twitter: </em></strong><a href="https://twitter.com/r3v3Ncl4W"><em>https://twitter.com/r3v3Ncl4W</em></a></blockquote><blockquote><strong><em>Instagram: </em></strong><a href="https://instagram.com/r3v3ncl4w.go"><em>https://instagram.com/r3v3ncl4w.go</em></a></blockquote><blockquote><strong><em>Blog page: </em></strong><a href="https://instagram.com/forensis_digitum"><em>https://instagram.com/forensis_digitum</em></a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5497816da50" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Governance: From Buzzword to Business Lifeline]]></title>
            <link>https://yashgorasiya.medium.com/ai-governance-from-buzzword-to-business-lifeline-9b0d546d5545?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b0d546d5545</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 22 Sep 2025 13:44:04 GMT</pubDate>
            <atom:updated>2025-09-22T13:44:04.538Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o87lOrIz2fiPYaeEQoQWeQ.png" /></figure><p>Hello, everyone. Welcome to yet another blog around AI Security.</p><p>After last week’s buzzing time at BSides Ahmedabad, we are back again at it.</p><p>Last time, we dug into <em>why humans still matter in AI security</em>. Remember the line? <strong>Machines scale. Humans decide.</strong> That post was all about balance — AI gives us speed and scale, but humans bring judgement, ethics, and accountability.</p><p>Naturally, that raises the next big question: if AI and humans are now working side by side, <em>who sets the rules to make sure the partnership works safely, fairly, and legally?</em></p><p>That’s where <strong>AI Governance</strong> comes in.</p><p>I know, the word “governance” sounds like a boardroom buzzword — the sort of thing that gets thrown into slide decks. But here’s the reality: governance is the <strong>difference between AI being a competitive advantage or a legal and reputational nightmare.</strong></p><p>Think of governance like the <strong>seatbelts and traffic lights of AI.</strong> Without it, things don’t just move fast — they crash. With it, businesses can innovate at speed, but also keep control, trust, and compliance.</p><p>And trust me, after digging through ICO rulings, government white papers, and real-world cases, I can confidently say: <strong>AI governance isn’t a “nice-to-have” anymore. It’s survival.</strong></p><h3>Why AI Governance Matters (More Than Ever)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/240/1*evg_-1CTCqYwHcMZkSzfzQ.jpeg" /></figure><p>AI governance, simply put, is <strong>rules + oversight + accountability.</strong></p><p>But why does it matter <em>right now</em>?</p><h4>1. AI adoption is exploding</h4><p>AI is no longer niche. McKinsey reports that <strong>60% of businesses now use AI in some form</strong> — whether in fraud detection, chatbots, or supply chain forecasting. Half the time, managers don’t even realise how much sensitive data these models process.</p><p>📌 Without governance, that’s like giving an intern your company credit card with no spending limit.</p><h4>2. Regulators are getting sharper teeth</h4><p>The <strong>ICO’s updated guidance (2023)</strong> made it clear: fairness, transparency, explainability, and accountability aren’t optional — they’re legal requirements. [<a href="https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/">ICO Guidance</a>]</p><p>And regulators are ready to bite. Just look at <strong>Meta’s €1.2 billion GDPR fine (2023)</strong> — not directly AI, but a crystal-clear reminder that sloppy data practices can cost eye-watering amounts.</p><p>📌 The lesson: regulators don’t care how shiny your AI tool is. If you can’t prove fairness and accountability, expect a knock on the door.</p><h4>3. Risks are already showing.</h4><p>IBM’s 2025 research revealed <strong>13% of organisations suffered breaches in their AI systems</strong> — and an unbelievable <strong>97% of those lacked proper AI access controls.</strong></p><p>📌 In plain terms: businesses built powerful AI systems… then forgot to lock the doors.</p><h3>What Happens Without Governance (Lessons From the Field)</h3><p>Governance gaps aren’t theory — they play out in real disasters. Let’s dig into a few and the lessons they teach.</p><h4>DeepMind &amp; NHS (2017)</h4><p>Google’s DeepMind struck a deal with London’s Royal Free NHS Trust to test its “Streams” health app. But in the process, it accessed <strong>1.6 million patient records</strong> without telling patients. The ICO later ruled this unlawful.</p><ul><li><strong>What went wrong:</strong> No transparency, no consent, no oversight.</li><li><strong>The lesson:</strong> Even if your goal is “improving healthcare,” you can’t skip governance. Patients must know where their data goes.</li></ul><p>This wasn’t a hack — it was a governance failure.</p><h4>COMPAS Algorithm (US Courts)</h4><p>Judges in the US used COMPAS, an AI risk assessment tool, to predict whether defendants would reoffend. Investigations found it unfairly flagged Black defendants as “high risk” more often than white defendants.</p><ul><li><strong>What went wrong:</strong> Biased training data, no audits, no accountability.</li><li><strong>The lesson:</strong> Without governance, AI simply amplifies existing biases at scale. What looked like “objective math” was actually discriminatory.</li></ul><p>This case became the poster child for why fairness audits aren’t optional.</p><h4>ChatGPT Data Leak (2023)</h4><p>A simple bug in OpenAI’s system exposed user chats and even payment details to strangers. The issue wasn’t malicious hacking — it was a lack of strict governance in deployment and testing.</p><ul><li><strong>What went wrong:</strong> Poor testing, weak access controls.</li><li><strong>The lesson:</strong> Even cutting-edge AI tools need boring-but-crucial governance steps: QA, security reviews, and access logging.</li></ul><p>Governance isn’t exciting — but it prevents these “oops” moments.</p><h4>Meta GDPR Fine (2023)</h4><p>Meta was fined <strong>€1.2 billion</strong> for mishandling data transfers under GDPR. While not AI-specific, it shows how regulators punish governance failures.</p><ul><li><strong>What went wrong:</strong> Poor compliance processes, no lawful basis for certain data uses.</li><li><strong>The lesson:</strong> Governance isn’t just about technology — it’s about building processes to stay compliant.</li></ul><h4>The Hard Costs</h4><p>The IBM <em>Cost of a Data Breach Report 2023</em> puts the average breach cost at <strong>$4.45 million</strong>. AI-driven breaches are even more expensive because they’re harder to untangle. [<a href="https://www.ibm.com/reports/data-breach">IBM</a>]</p><ul><li><strong>What went wrong:</strong> Lack of preparedness, no governance layers.</li><li><strong>The lesson:</strong> It’s always cheaper to govern upfront than to pay for breaches later.</li></ul><p>📌 Bottom line: Governance gaps don’t just cause fines. They break trust, amplify bias, and cost millions.</p><h3>Humans + AI: Who’s Accountable?</h3><p>This is the most uncomfortable part of AI. If a model makes a bad call, <em>who answers for it?</em></p><p>Imagine this: your AI tool rejects a customer’s mortgage. The customer complains, and regulators come knocking. Who takes the heat?</p><ul><li>The <strong>developer</strong>? “We just wrote the code.”</li><li>The <strong>vendor</strong>? “We just sold the product.”</li><li>The <strong>company using it</strong>? Bingo.</li></ul><p>The ICO is crystal clear: <strong>accountability lies with the organisation deploying AI.</strong> If you’re using it, you own the outcomes — good or bad.</p><h4>Why This Matters</h4><ul><li><strong>Legal liability:</strong> You can’t say, “the algorithm did it.” Regulators don’t buy that excuse.</li><li><strong>Reputational risk:</strong> Customers don’t care whose code it was — they care who denied them service.</li><li><strong>Operational reality:</strong> Your employees are the ones fielding complaints when AI goes wrong.</li></ul><h4>UK Guidance on Accountability</h4><ul><li><strong>ICO Guidance (2023):</strong> Organisations must document who owns which AI system, who audits it, and how risks are managed. [<a href="https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/">ICO</a>]</li><li><strong>UK White Paper (2023):</strong> Encourages “pro-innovation” AI but makes accountability non-negotiable — transparency and fairness must be provable. [<a href="https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper">Gov UK</a>]</li></ul><p>📌 Simple truth: If you use AI in your business, you can’t outsource accountability. You own it.</p><h3>The UK Lens</h3><p>The UK’s position is tricky: not in the EU, but not ignoring it either. That creates a unique governance environment.</p><h4>1. ICO Guidance on AI &amp; Data Protection</h4><p>Updated in 2023, it lays down five essentials:</p><ul><li>Fairness</li><li>Transparency</li><li>Accountability</li><li>Data minimisation</li><li>Explainability</li></ul><p>This means you need to be able to tell <em>any user</em>: “Here’s why AI made this decision, here’s what data it used.”</p><h4>2. NCSC + CISA Guidelines (2023)</h4><p>The UK’s National Cyber Security Centre teamed up with US CISA to publish <em>Guidelines for Secure AI System Development</em>. It’s basically a checklist for:</p><ul><li>Securing AI pipelines.</li><li>Guarding against adversarial inputs.</li><li>Ensuring monitoring throughout the lifecycle.</li></ul><p>For UK firms, this isn’t “nice advice” — it’s a best-practice baseline.</p><h4>3. EU AI Act (2024)</h4><p>Even post-Brexit, UK companies that operate in Europe will have to comply. High-risk AI (like healthcare, finance, law enforcement) will need strict governance: bias audits, human oversight, and explainability mechanisms.</p><p>📌 For UK businesses, this means governance isn’t just local — it’s global.</p><h4>4. Public Trust</h4><p>A YouGov poll showed that <strong>only 28% of UK adults trust companies to use AI responsibly.</strong> Governance isn’t just about regulators — it’s about convincing customers you won’t abuse their data.</p><h3>Practical Playbook: How to Build Governance That Works</h3><p>Theory is fine. But how do you <em>actually</em> govern AI?</p><h4>This Week (Quick Wins)</h4><ul><li><strong>AI Register:</strong> Log every AI system, what it does, what data it touches, and who owns it.</li><li><strong>AI Principles Page:</strong> Publish a short transparency note for staff and customers.</li><li><strong>Access Audit:</strong> Pick one AI tool and review who has access and whether it’s logged.</li></ul><h4>This Month (Operational Fixes)</h4><ul><li><strong>DPIA:</strong> Run a Data Protection Impact Assessment on a high-risk AI tool.</li><li><strong>Assign Roles:</strong> Appoint a Model Owner, a Risk Officer, and a Legal sign-off.</li><li><strong>Explainability Test:</strong> Can you explain AI decisions in plain terms? If not → fix.</li></ul><h4>This Quarter (Deep Governance)</h4><ul><li><strong>Bias &amp; Drift Audit:</strong> Check if models treat certain groups unfairly, retrain if needed.</li><li><strong>AI Failure Drill:</strong> Simulate an AI gone wrong — who steps in, who explains to customers, who reports to regulators?</li><li><strong>Build Into SDLC:</strong> Make governance part of development, not an afterthought.</li></ul><h4>Scenario in action</h4><p>A bank’s AI fraud system flags late-night transactions as “suspicious.” With governance:</p><ol><li>AI Register shows the system is owned by Risk Ops.</li><li>Risk Officer checks and confirms: these are regular transactions by night-shift workers.</li><li>Bias audit reveals a pattern → retraining scheduled.</li></ol><p>No panic. No reputational damage. Just governance doing its job.</p><h3>My POV</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lSrcqmDmg_WUXrONTOfUsA.png" /></figure><p>Here’s where I land after weeks of digging into this.</p><p>AI without governance is like giving a teenager a Ferrari with no licence. Yes, it can go fast. But when it crashes — and it will — who gets blamed? The owner.</p><p>For me, AI governance is:</p><ul><li><strong>Not red tape.</strong> It’s a seatbelt. You don’t put one on because you like it. You put it on because you want to get home safe.</li><li><strong>Not optional.</strong> Regulators will force it sooner or later. Smart companies adopt it early, on their terms.</li><li><strong>Not just compliance.</strong> Good governance builds trust — with customers, staff, and investors.</li></ul><p>If I were talking to a board tomorrow, I’d say this:</p><ul><li>Don’t measure success by “how many AI tools we deployed.”</li><li>Measure success by “how many governance surprises we avoided.”</li></ul><p>Because every surprise avoided = millions saved + trust intact.</p><p>👉 My bottom line: <strong>Machines scale. Humans decide. Governance makes sure humans stay accountable and customers stay protected.</strong></p><h3>Final Takeaway</h3><p>AI without governance is like building a skyscraper without fire exits — impressive until something goes wrong.</p><p>The winners of the AI era won’t just deploy models. They’ll govern them. Making AI fair, transparent, and accountable isn’t just about compliance — it’s about building the trust that keeps customers and regulators on your side.</p><h3>❤️ Thanks for Reading</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ijkBiwWhoGJ8D0j30NVCaQ.png" /></figure><p>Thanks for sticking with me till here! 🙏</p><p>Please don’t forget to leave a review or share your thoughts in the comments <strong>🥳🙏🫶 </strong>.</p><p>And hey, if you found this helpful, <strong>subscribe</strong> to stay in the loop for future blogs.</p><p><strong>Together, let’s build a safer and more secure digital world! 💪🤗👨💻</strong></p><blockquote><strong><em>Get connected with me here :</em></strong></blockquote><blockquote><strong>LinkedIn: </strong><a href="https://www.linkedin.com/in/yash-gorasiya/">https://www.linkedin.com/in/yash-gorasiya/</a></blockquote><blockquote><strong>Twitter: </strong><a href="https://twitter.com/r3v3Ncl4W">https://twitter.com/r3v3Ncl4W</a></blockquote><blockquote><strong>Instagram: </strong><a href="https://instagram.com/r3v3ncl4w.go">https://instagram.com/r3v3ncl4w.go</a></blockquote><blockquote><strong>Blog page: </strong><a href="https://instagram.com/forensis_digitum">https://instagram.com/forensis_digitum</a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b0d546d5545" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Human Factor Isn’t Dead: Why People Still Matter in AI Security]]></title>
            <link>https://yashgorasiya.medium.com/the-human-factor-isnt-dead-why-people-still-matter-in-ai-security-d62e4353166f?source=rss-835e0d12af1------2</link>
            <guid isPermaLink="false">https://medium.com/p/d62e4353166f</guid>
            <category><![CDATA[people]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[human-factors]]></category>
            <dc:creator><![CDATA[Yash Gorasiya]]></dc:creator>
            <pubDate>Mon, 08 Sep 2025 08:31:40 GMT</pubDate>
            <atom:updated>2025-09-08T08:31:40.823Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1BjRzBp_Ye5NHWUp9-iHnA.png" /></figure><p>Hey folks 👋, glad to have you here again. Over the last few weeks, I’ve been diving into how AI is reshaping security — sometimes making us feel safer, sometimes making us worry about Big Brother vibes.</p><p>This week, I want to talk about something I’ve been thinking about a lot while reading reports, attending webinars, and even swapping notes with a few people in the industry:</p><p>👉 <em>Can AI really replace humans in cybersecurity? Or are people still the most important part of the system?</em></p><p>At first glance, it’s tempting to think machines might take over completely. After all, they don’t get tired, don’t miss things buried in logs, and don’t need coffee breaks. But as I’ve gone deeper, I keep running into examples and case studies that show something different: <strong>AI may be the engine, but humans are still the drivers.</strong></p><h3>Why the “AI replaces humans” story is misleading</h3><p>Imagine you’re in a self-driving car with no steering wheel. Sounds cool until the car faces unusual roadworks or a weird weather condition it didn’t expect.</p><p>In security, “weird” is normal.</p><p>Attackers invent new tricks daily. AI is fantastic at scale and pattern-matching, but when the unexpected happens, we still need judgment, ethics, intuition and common sense — human skills that machines don’t have.</p><p>Two stats really jumped out at me when I started researching this:</p><ul><li>The <strong>average cost of a data breach</strong> globally is about <strong>$4.45 million</strong> now. That’s not just a number — it’s salaries lost, projects delayed, customers leaving, and fines piling up. For many mid-sized businesses, one big breach can literally mean the end.</li><li>And here’s the kicker: according to Verizon’s DBIR, <strong>74% of those breaches involve a human element</strong> — someone clicking a dodgy link, reusing a password, or being tricked by a scam email.</li></ul><p>Now, you could read that and think: “So humans are the problem.” But the way I see it, it’s the opposite. If people are central to how breaches happen, then people are also central to how breaches get stopped.</p><p>📌 <strong>Think about it this way:</strong> Fire alarms don’t put out fires. They just warn you. You still need firefighters — trained humans — to respond and make decisions on the ground. AI is like the alarm: powerful at detection. But the people are the ones who ultimately put the fire out.</p><h3>What AI Does Brilliantly (and Why That’s Not the Whole Story)</h3><p>Let’s give AI its due credit — because honestly, some of the things it does are mind-blowing.</p><h4>1. Speed and Scale That Humans Can’t Match</h4><p>AI can chew through millions of log entries in seconds and point out anomalies that would take a human team days to even notice.</p><ul><li>Imagine sorting through every single CCTV camera in London at once, looking for someone in a red jacket. A human team would be exhausted in hours. AI? It’ll do it before you finish your tea.</li></ul><p>That’s why big organisations lean so hard on AI for threat detection — the sheer volume of digital “noise” is impossible for humans to handle alone.</p><h4>2. Predictive Defence (Like a Weather Forecast for Attacks)</h4><p>AI doesn’t just spot what already happened. It can also compare global attack patterns and give early warnings.</p><ul><li>For example, if hospitals in Germany are getting hit by a new ransomware strain, AI tools monitoring international feeds can alert NHS systems to brace for impact.</li></ul><p>It’s like the Met Office predicting a storm front. You might not see the clouds yet, but you can start packing the sandbags.</p><h4>3. Automation of the Boring Stuff</h4><p>One of the biggest drains in cybersecurity is repetitive tasks: resetting passwords, scanning for vulnerabilities, and updating patches.</p><ul><li>AI takes those chores off the table, freeing up human analysts to focus on the weird, complex stuff that actually needs judgment.</li></ul><p>It’s the digital equivalent of having a dishwasher. You <em>can</em> do the washing-up yourself, but wouldn’t you rather spend that time cooking the actual meal?</p><h4>But here’s the key: these are <strong>tools, not replacements.</strong></h4><p>AI can flag anomalies, predict risks, and handle grunt work. But when it comes to making sense of what those anomalies mean, deciding how to respond, or communicating with people — that’s still very much human territory.</p><h3>Where AI Falls Short (and Why People Still Matter)</h3><p>For all the good AI brings, there are places where it just can’t cut it. These are the gaps that keep reminding me: humans aren’t going anywhere in cybersecurity.</p><h4>1. AI Doesn’t Understand Intent</h4><p>AI is brilliant at spotting patterns, but it doesn’t know <em>why</em> something is happening.</p><ul><li>Example: If John from Finance logs in at 2 a.m. from Spain, AI might throw a red flag.</li><li>But maybe John’s at a conference in Barcelona and just uploading the quarterly report before bed.</li></ul><p>A machine sees “unusual login at an odd time.”<br>A human sees “John is travelling, and this makes sense.”</p><p>📌 Without that context, AI can lock people out or flood the SOC with false alarms.</p><h4>2. AI Can’t Handle Manipulation</h4><p>Hackers know how to play people, not just systems. And that’s a huge blind spot for machines.</p><ul><li>Real case: In 2019, scammers used an <strong>AI-cloned voice of a CEO</strong> to trick an employee into wiring <strong>£200,000</strong> to their account (<a href="https://www.bbc.com/news/technology-48908736">BBC</a>).</li><li>The request sounded urgent and believable — AI tools didn’t block it, because technically, nothing “looked” wrong in the transaction.</li></ul><p>It took human suspicion and verification afterwards to catch it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cPeKGD5ytxotIkJsf60yuA.jpeg" /></figure><p>Think <em>Inception</em>. Hackers don’t just attack your systems — they attack your perception of reality. AI isn’t built for that.</p><h4>3. AI Can Be Biased</h4><p>Machines are only as good as the data they’re trained on. If the training set is skewed, so are the results.</p><ul><li>Example: Imagine training an AI system mostly on office workers’ 9–5 data.</li><li>Now throw in NHS night-shift doctors logging in at 3 a.m.</li><li>The AI flags them as suspicious every time, simply because they don’t “fit the norm.”</li></ul><p>This isn’t security. It’s bias. And it wastes everyone’s time. Humans are needed to spot and correct these blind spots.</p><h4>4. AI Doesn’t Do Ethics or Trade-offs</h4><p>Sometimes security decisions aren’t just technical — they’re moral or strategic.</p><ul><li>Example: A ransomware attack hits a hospital. AI says: “Shut everything down to stop the spread.”</li><li>But humans have to ask: “If we shut it down, does patient care stop? Do we risk lives?”</li></ul><p>That’s not something you want left to an algorithm. It needs people weighing business impact, ethics, and human consequences.</p><h3>5. Attackers Use AI Too</h3><p>Here’s the part that makes me sit up straight: the other side has AI as well.</p><ul><li>Hackers are using it to craft hyper-realistic phishing emails, generate deepfake videos, and even build malware that adapts on the fly.</li><li>The <strong>Synnovis ransomware attack in London (2024)</strong> disrupted NHS pathology labs and leaked nearly <em>1 million patient records</em>. Reports suggested attackers used AI-enhanced tools to evade defences.</li></ul><p>It’s not just humans vs. machines anymore. It’s machines vs. machines — with humans caught in the middle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/1*R1C-ixz9PrgvRtx2lomqjQ.jpeg" /></figure><p>It’s like Iron Man fighting Ultron. Both sides have the tech, but at the end of the day, Tony Stark’s human decisions make the difference.</p><h4>The big lesson?</h4><p>AI is incredible at <strong>spotting signals</strong>. But when it comes to <strong>making sense of them, handling manipulation, balancing trade-offs, or staying ahead of attackers</strong>, humans are still the difference between chaos and control.</p><h3>Why This Matters in the UK</h3><p>Now, you might think all of this is just Silicon Valley talk, but honestly — the UK is right in the thick of it.</p><ul><li>The UK Government’s <strong>Cyber Security Breaches Survey 2023</strong> found that most medium and large businesses had identified a cyber incident in the last year. And here’s the kicker: phishing is still the number one attack, meaning people are still the prime targets. (<a href="https://www.gov.uk/government/statistics/cyber-security-breaches-survey-2023">UK Gov</a>)</li><li>Regulators aren’t sitting back either. Remember when <strong>Meta was fined €1.2 billion in 2023</strong> for breaching GDPR? That wasn’t some slap on the wrist — it was a signal that if you mishandle data, regulators <em>will</em> make an example out of you.</li><li>And the <strong>ICO’s guidance on AI and data protection</strong> makes it crystal clear: if you’re using AI in ways that touch personal data, you need transparency, fairness, and accountability built in. (<a href="https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/explaining-decisions-made-with-artificial-intelligence/">ICO</a>)</li></ul><p>So for UK businesses, this isn’t about shiny tools. It’s about trust, regulation, and staying in business.</p><h3>Practical Playbook (What to Do About It)</h3><p>Alright, let’s get practical. Because talking about risks is fine, but what can you actually do next week, next month, next quarter? Here’s how I’d break it down.</p><h4>This Week (Quick Wins)</h4><ul><li><strong>Write a one-page explainer</strong> for staff: say what your AI tools monitor, what they don’t, and why. Please keep it simple.</li><li><strong>Change one risky automation</strong> from “auto-block” to “raise-to-human.” For example: high-value payment requests.</li><li><strong>Add an AI column to your asset register</strong> so you know which systems are using machine learning and on what data.</li></ul><h4>This Month (Operational Fixes)</h4><ul><li><strong>Set up an out-of-band verification process</strong>: finance staff must confirm big payments with a phone call to a pre-registered number.</li><li><strong>Run a tabletop exercise</strong>: simulate a deepfake CEO voice call demanding a transfer. Practice saying, “Hang on, let me call you back.”</li><li><strong>Check your false positives</strong>: are certain groups (like remote workers or night staff) getting flagged more often? That’s bias. Tune your models.</li></ul><h4>This Quarter (Governance and Oversight)</h4><ul><li><strong>Form a small AI governance group</strong>: one data protection officer, one security lead, and someone from ops. They own audits and risk reviews.</li><li><strong>Audit your models for drift</strong>: if they were trained six months ago, are they still accurate today?</li><li><strong>Do an “AI failure mode” tabletop</strong>: what happens if the AI flags the wrong thing, or misses an attack? Practice the human backstops.</li></ul><h4><strong>A quick scenario to picture this:</strong></h4><p>Imagine payroll gets a Slack DM from “the CFO” asking to approve a payment. AI flags it as unusual. Instead of auto-blocking, it escalates to a SOC analyst, who then calls the CFO’s <em>real</em> number. The CFO, sitting on a delayed flight, says, “Nope, not me.” Payment blocked. Fraud avoided.</p><p>That’s not futuristic. That’s what happens when AI and humans work <em>together</em>.</p><h3>My POV</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QmaSrkgMfqbFC4slnvzeQg.png" /></figure><p>Here’s where I land after digging into this topic.</p><p>AI is powerful, no doubt. It’s like having a super-fast assistant who can sift through mountains of data in seconds. But left alone, it’s blind to context, bias, ethics, and trust. That’s where people come in.</p><p>If I were advising a leader, I’d say:</p><ul><li>Let AI do the heavy lifting.</li><li>Let people handle the judgment calls.</li><li>Don’t sell AI as a “replacement” for people — sell it as “extra eyes and faster legs.”</li></ul><p>Because when things go wrong (and they always do), accountability doesn’t sit with the algorithm. It sits with the humans who built, governed, and deployed it.</p><p>👉 My bottom line: <strong>Machines scale. Humans decide.</strong> The strongest cybersecurity strategies will always be the ones where they work hand in hand.</p><h3>Final takeaway</h3><p>AI is a powerful force multiplier for security. But it’s a tool, not a replacement for judgment, ethics, and context. The organisations that win will be the ones who design AI to spot signals and humans to give meaning to them — fast, fair, and with trust at the centre.</p><h3>❤️ Thanks for Reading</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/0*PxToFW6sFIp0Enww.png" /></figure><p>Thanks for sticking with me till here! 🙏</p><p>Please don’t forget to leave a review or share your thoughts in the comments <strong>🥳🙏🫶 </strong>.</p><p>And hey, if you found this helpful, <strong>subscribe</strong> to stay in the loop for future blogs.</p><p><strong>Together, let’s build a safer and more secure digital world! 💪🤗👨‍💻</strong></p><blockquote><strong>Get connected with me here :</strong></blockquote><blockquote><strong><em>LinkedIn: </em></strong><a href="https://www.linkedin.com/in/yash-gorasiya/"><em>https://www.linkedin.com/in/yash-gorasiya/</em></a></blockquote><blockquote><strong><em>Twitter: </em></strong><a href="https://twitter.com/r3v3Ncl4W"><em>https://twitter.com/r3v3Ncl4W</em></a></blockquote><blockquote><strong><em>Instagram: </em></strong><a href="https://instagram.com/r3v3ncl4w.go"><em>https://instagram.com/r3v3ncl4w.go</em></a></blockquote><blockquote><strong><em>Blog page: </em></strong><a href="https://instagram.com/forensis_digitum"><em>https://instagram.com/forensis_digitum</em></a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d62e4353166f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>