<?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 Viren Choudhari on Medium]]></title>
        <description><![CDATA[Stories by Viren Choudhari on Medium]]></description>
        <link>https://medium.com/@virenchoudhari?source=rss-b8694a2687db------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*fCBSZUzv4XNLJwDt.jpg</url>
            <title>Stories by Viren Choudhari on Medium</title>
            <link>https://medium.com/@virenchoudhari?source=rss-b8694a2687db------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 11:11:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@virenchoudhari/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 Asset Inventory: What Counts as an AI Asset?]]></title>
            <link>https://virenchoudhari.medium.com/ai-asset-inventory-what-counts-as-an-ai-asset-8e2ab5c7c4fd?source=rss-b8694a2687db------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e2ab5c7c4fd</guid>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[nhis]]></category>
            <category><![CDATA[grc]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Viren Choudhari]]></dc:creator>
            <pubDate>Fri, 16 Jan 2026 03:52:23 GMT</pubDate>
            <atom:updated>2026-01-16T03:52:23.547Z</atom:updated>
            <content:encoded><![CDATA[<p>Security teams are used to inventorying laptops, servers, containers, SaaS apps, and APIs. But AI changes the shape of “what matters” — and the biggest gap we see isn’t model security. It’s simply not knowing <em>where AI exists</em> in the enterprise.</p><p>An AI asset inventory is the foundation for every other control: risk prioritization, policy enforcement, incident response, compliance reporting, and vendor governance. If you don’t define what counts as an AI asset, you’ll end up with a partial list — and a false sense of coverage.</p><p>This post gives you a practical definition of AI assets, a taxonomy you can use immediately, and a workflow to build an inventory that stays current.</p><p><strong>Why an AI asset inventory is different from a software inventory</strong></p><p>Traditional inventory approaches assume:</p><ul><li>assets are owned by IT,</li><li>run on known infrastructure,</li><li>have a clear boundary (server/app/service),</li><li>and change slowly.</li></ul><p>AI assets often break those assumptions:</p><ul><li><strong>They’re embedded</strong> inside apps you already have (“AI-powered search,” “chat assistant,” “smart triage”).</li><li><strong>They’re composable</strong> (a single user workflow may chain an LLM API, a vector database, a prompt template, a plugin, and an agent runner).</li><li><strong>They change fast</strong> (prompt updates, model swaps, new tools, new data sources).</li><li><strong>They pull in data</strong> (and data risk is often the real risk).</li></ul><p>So the inventory must capture more than “what model are we using?” It must capture the components <em>around</em> the model.</p><p><strong>A simple definition: what counts as an AI asset?</strong></p><p><strong>An AI asset is any system component that creates, transforms, ranks, or acts on information using machine learning or generative AI — especially when it:</strong></p><ul><li>touches enterprise data,</li><li>influences decisions,</li><li>produces user-facing content,</li><li>or can execute actions (directly or indirectly).</li></ul><p>If it can <em>change an outcome</em> (what users see, what gets approved, what gets sent, what gets deleted), it belongs in the inventory.</p><p><strong>The AI asset taxonomy (use this as your inventory categories)</strong></p><p>Think in layers. Most orgs focus on the model layer and miss everything else.</p><p><strong>1) AI-enabled applications (the “you already bought it” layer)</strong></p><p>These are SaaS products or internal apps that include AI features:</p><ul><li>CRM copilots, meeting assistants, ticket triage</li><li>code assistants in IDEs</li><li>AI-driven analytics in BI tools</li><li>email/security products with “AI-based detection” or “GenAI summaries”</li></ul><p><strong>Inventory rule:</strong> If the vendor markets an AI feature, treat the feature as an AI asset — even if you never deployed a “model.”</p><p><strong>2) Models (hosted, fine-tuned, local)</strong></p><p>Models are still assets — but they’re only one category:</p><ul><li>third-party hosted LLMs (API access)</li><li>open-source LLMs deployed on your infra</li><li>fine-tuned models (your weights)</li><li>classical ML models used for scoring/classification</li><li>model variants (prod vs staging vs experimental)</li></ul><p><strong>Inventory rule:</strong> Track <em>version</em>, <em>hosting</em>, and <em>who can call it</em>.</p><p><strong>3) AI services and endpoints</strong></p><p>Anything exposing AI capability as a callable service:</p><ul><li>inference endpoints</li><li>gateways/proxies</li><li>model routers</li><li>safety/filter services</li><li>embedding services</li></ul><p><strong>Inventory rule:</strong> If it has a URL, auth, and logs, it’s an asset.</p><p><strong>4) Prompts, prompt templates, and system instructions</strong></p><p>Prompts are executable business logic:</p><ul><li>system prompts (“you are a financial assistant…”)</li><li>routing prompts (“if user asks X, do Y”)</li><li>prompt templates embedded in apps</li><li>prompt repositories used by multiple teams</li></ul><p><strong>Inventory rule:</strong> If changing a prompt changes behavior, treat it like code: version it, own it, review it.</p><p><strong>5) Agents and tools (action surfaces)</strong></p><p>Agents introduce “do things” risk:</p><ul><li>tool-using agents (email sender, ticket closer, code PR creator)</li><li>browser automation</li><li>RPA integrations</li><li>internal tools (HR lookup, customer database query, refund tool)</li></ul><p><strong>Inventory rule:</strong> If the AI can trigger an action (even with approval), it belongs in the inventory with a clear permission boundary.</p><p><strong>6) Data connectors and retrieval systems (RAG layer)</strong></p><p>This is where many real-world incidents happen:</p><ul><li>knowledge base connectors (SharePoint, Confluence, Google Drive)</li><li>data warehouse connectors</li><li>vector databases / embedding indices</li><li>retrieval pipelines (chunking, indexing, refresh jobs)</li><li>document parsers and OCR components</li></ul><p><strong>Inventory rule:</strong> If it can expose sensitive data to a model or user, it’s an asset.</p><p><strong>7) Training and evaluation pipelines (MLops/LLMOps layer)</strong></p><p>Not every org trains models, but many tune, evaluate, or run tests:</p><ul><li>fine-tuning jobs</li><li>evaluation harnesses and benchmarks</li><li>red-teaming suites</li><li>prompt testing pipelines</li><li>guardrail configuration and policy engines</li></ul><p><strong>Inventory rule:</strong> Anything that changes model behavior or measures it belongs in scope.</p><p><strong>8) Datasets (often the highest-risk “AI asset”)</strong></p><p>AI-related datasets include:</p><ul><li>training datasets</li><li>retrieval corpora</li><li>prompt logs / conversation logs</li><li>feedback datasets (thumbs up/down, labels)</li><li>synthetic data generation outputs</li></ul><p><strong>Inventory rule:</strong> If it’s used to teach, retrieve, or improve AI behavior, treat it as an AI dataset asset with classification and retention.</p><p><strong>9) Identity, access, and secrets specific to AI</strong></p><p>This category is frequently missed:</p><ul><li>API keys for model providers</li><li>service accounts used by agents/tools</li><li>OAuth tokens for connectors</li><li>credentials embedded in pipelines</li></ul><p><strong>Inventory rule:</strong> If compromise would allow model access or data exfiltration, it’s an AI asset.</p><p><strong>10) Third-party AI providers and embedded AI supply chain</strong></p><p>Even if you didn’t “buy an AI platform,” you still have AI exposure through vendors:</p><ul><li>SaaS vendors using AI under the hood</li><li>call-center tools that process transcripts</li><li>document processing vendors using ML</li></ul><p><strong>Inventory rule:</strong> If they process your data with AI, record them as AI vendors with data types + purpose.</p><p><strong>Common blind spots (where your inventory will be incomplete)</strong></p><ol><li><strong>“Shadow AI” usage</strong>: employees using public AI tools with work data</li><li><strong>AI features inside existing SaaS</strong>: enabled by default, not tracked as a “new system”</li><li><strong>Prompts living in code</strong>: scattered across repos and configs</li><li><strong>Connectors</strong>: the “pipes” to internal documents and systems</li><li><strong>Evaluation data</strong>: logs used to improve prompts/models contain sensitive data</li><li><strong>Agent tools</strong>: a small tool permission mistake becomes a major incident path</li></ol><p><strong>What to record for each AI asset (minimum viable fields)</strong></p><p>You want just enough structure to be useful <em>and</em> sustainable.</p><p><strong>Minimum fields</strong></p><ul><li>Asset name + type (from taxonomy above)</li><li>Business owner and technical owner</li><li>Environment (prod/stage/dev)</li><li>Hosting (SaaS / cloud / on-prem) + region</li><li>Data classification touched (public/internal/confidential/regulated)</li><li>Entry points (users, apps, APIs)</li><li>Dependencies (models, vector DB, connectors, tools)</li><li>Authentication method (SSO, API key, service account)</li><li>Logging and monitoring status (yes/no + where)</li><li>Guardrails in place (policy, filters, approval gates)</li><li>Change control (how prompts/models are updated)</li><li>Incident contact + runbook link</li></ul><p><strong>Optional but powerful</strong></p><ul><li>“Worst-case action” the asset can trigger (send email, modify records, approve transactions)</li><li>Known failure modes (hallucination risk, data leakage risk)</li><li>Measured quality + drift signals</li><li>Compliance mapping (e.g., internal AI policy controls)</li></ul><p><strong>How to build the inventory in 30 days (a practical workflow)</strong></p><p><strong>Step 1: Start with procurement + SaaS catalogs</strong></p><p>Pull from:</p><ul><li>app catalog / SSO portal (anything with “copilot,” “assistant,” “AI”)</li><li>vendor questionnaires / DPAs mentioning ML/GenAI</li><li>expense tools (subscriptions to AI tools)</li></ul><p>Output: <strong>AI-enabled SaaS list</strong>.</p><p><strong>Step 2: Scan the engineering surface area</strong></p><p>Target:</p><ul><li>repos with prompt strings, “openai/anthropic/bedrock/vertex” SDK usage</li><li>API gateways or inference endpoints</li><li>vector DB usage (Pinecone, Weaviate, pgvector, etc.)</li><li>“agent/tool” frameworks and tool definitions</li></ul><p>Output: <strong>AI in code</strong> (models + prompts + endpoints + tools).</p><p><strong>Step 3: Map data connectors and retrieval sources</strong></p><p>For each AI app/service ask:</p><ul><li>what data sources are connected?</li><li>what’s indexed?</li><li>how often is it refreshed?</li><li>who can query it?</li><li>how do we delete/expire data?</li></ul><p>Output: <strong>RAG + connector map</strong>.</p><p><strong>Step 4: Capture identity + secrets</strong></p><p>Work with IAM/security engineering:</p><ul><li>where are AI API keys stored?</li><li>what service accounts exist for AI agents?</li><li>what scopes do connectors have?</li></ul><p>Output: <strong>AI credential inventory</strong>.</p><p><strong>Step 5: Put it in one place and keep it alive</strong></p><p>Pick a system of record:</p><ul><li>CMDB + custom fields</li><li>security asset inventory tool</li><li>a dedicated “AI Assets” register (even a database-backed form)</li></ul><p>Then define:</p><ul><li>who updates it (RACI)</li><li>update triggers (new AI feature, new connector, new model, new tool)</li><li>a lightweight quarterly review</li></ul><p><strong>A simple test: do we have an AI asset, yes or no?</strong></p><p>Use these yes/no questions:</p><ul><li>Does it use ML/GenAI to generate, rank, classify, summarize, or decide?</li><li>Does it access enterprise data (directly or via retrieval)?</li><li>Can it influence a user decision or business outcome?</li><li>Can it take actions (even with approval)?</li><li>Does it rely on a third-party AI provider?</li></ul><p>If <strong>yes</strong> to any: it belongs in the AI asset inventory.</p><p><strong>What “good” looks like</strong></p><p>A mature AI inventory lets you answer quickly:</p><ul><li>“Which AI assets touch regulated data?”</li><li>“Which assets can send emails or modify records?”</li><li>“Where are prompts stored and who can change them?”</li><li>“Which connectors expose sensitive repositories?”</li><li>“Which model endpoints have weak auth/logging?”</li><li>“If a provider has an incident, what depends on it?”</li></ul><p>That’s not just governance. That’s operational security.</p><p><strong>Closing: inventory first, everything else second</strong></p><p>AI security programs often start with model safety. But in practice, most incidents route through <strong>data access, identity/secrets, and tool permissions</strong> — all of which you only see if your inventory is complete.</p><p>Start with a clear definition, use a layered taxonomy, and treat prompts, connectors, and agent tools as first-class assets. Once you can list your AI assets, you can finally prioritize and reduce AI risk with the same discipline you apply to endpoints, cloud, and applications.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e2ab5c7c4fd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Used a Neo4j Attack Graph to Uncover a Hidden Lateral Movement Chain (A → B → C)]]></title>
            <link>https://virenchoudhari.medium.com/how-i-used-a-neo4j-attack-graph-to-uncover-a-hidden-lateral-movement-chain-a-b-c-5bc9ce2830ae?source=rss-b8694a2687db------2</link>
            <guid isPermaLink="false">https://medium.com/p/5bc9ce2830ae</guid>
            <category><![CDATA[neo4j]]></category>
            <category><![CDATA[attack-path]]></category>
            <dc:creator><![CDATA[Viren Choudhari]]></dc:creator>
            <pubDate>Sat, 29 Nov 2025 17:09:16 GMT</pubDate>
            <atom:updated>2025-11-29T17:09:16.794Z</atom:updated>
            <content:encoded><![CDATA[<h3>Most incident investigations stop when alerts stop.<br> But attackers don’t.</h3><p>In this write-up, I’ll show how I used an <strong>attack graph built in Neo4j</strong> to reconstruct a real multi-host lateral movement story across three machines — without packet captures, without EDR alerts, and without full logs.</p><p>The final result exposed a <strong>clean attack chain</strong> that looked like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WGY27pxSc20yk-YI2Gp-0Q.png" /></figure><blockquote><strong><em>NODE01 → SRV01 → DC101</em></strong><em><br> (PowerShell Remoting → NTLM lateral movement)</em></blockquote><p>Everything below comes from identity and network data alone.</p><h3>The Environment</h3><p>I worked with data collected from three Windows systems:</p><ul><li><strong>NODE01</strong> — a user workstation</li><li><strong>SRV01</strong> — a Windows Server acting as a pivot point</li><li><strong>DC101</strong> — Domain Controller</li></ul><p>Each host was ingested into Neo4j with:</p><ul><li>Processes</li><li>Users and profiles</li><li>Network connections</li><li>Ports</li><li>Domain membership</li></ul><p>Rather than looking at this as raw logs, I transformed it into a <strong>graph</strong>.</p><p>In Neo4j:</p><ul><li>Hosts became nodes</li><li>Users became nodes</li><li>Processes became nodes</li><li>Ports became nodes</li><li>Relationships tied everything together (owns, runs, exposes, connects)</li></ul><p>This changed the problem from <em>log investigation</em> to <em>path discovery</em>.</p><h3>Step 1 — The breach started on NODE01</h3><p>On <strong>NODE01</strong>, one process immediately stood out:</p><ul><li>powershell.exe</li><li>Owned by: NODE01\admin</li></ul><p>This is important because:</p><ul><li>PowerShell itself isn’t malicious.</li><li>But <strong>PowerShell running as Administrator</strong> often is the pivot point.</li></ul><p>In the graph, I could clearly see:</p><pre>NODE01 → powershell → NODE01\admin</pre><p>This told me:</p><blockquote><em>Whoever ran this had </em><strong><em>administrator-level control</em></strong><em> of the first machine.</em></blockquote><p>This is where the attack began.</p><h3>Step 2 — PowerShell Remoting to SRV01 (Confirmed)</h3><p>From NODE01, I then examined outbound connections.</p><p>One port was impossible to ignore:</p><ul><li><strong>5985 — WinRM (PowerShell Remoting)</strong></li></ul><p>The graph showed:</p><pre>NODE01 → Port 5985 → SRV01</pre><p>This isn’t just “network noise.”</p><p>WinRM on port 5985 is specifically used for:</p><ul><li>PowerShell remoting</li><li>Remote administration</li><li>Script execution across hosts</li></ul><p>And the owning process?</p><p>powershell.exe</p><p>And the account?</p><p>NODE01\admin</p><p>This gave me a confirmed lateral movement:</p><blockquote><em>The attacker opened a remote PowerShell session on SRV01 from NODE01.</em></blockquote><p>This wasn’t a guess.<br> This was architecture-level proof.</p><h3>Step 3 — SRV01 became the staging host</h3><p>On <strong>SRV01</strong>, identity signals exploded:</p><ul><li>Multiple logons</li><li>Multiple users</li><li>NTLM authentication</li><li>Administrator account present</li><li>Remote services running</li></ul><p>Ports exposed:</p><ul><li>5985 (WinRM)</li><li>445 (SMB)</li><li>135 (RPC)</li></ul><p>SRV01 looked exactly like what attackers use as a <strong>staging server</strong>:</p><ul><li>Close to the domain controller</li><li>Has credentials</li><li>Has services open</li><li>Has access in both directions</li></ul><p>At this point, the graph told me:</p><blockquote><em>SRV01 wasn’t just compromised — <br> it was being </em>used<em>.</em></blockquote><h3>Step 4 — The quiet hop to DC101</h3><p>This is where things get interesting.</p><p>There was no direct packet-level record from SRV01 → DC101.</p><p>But something else was visible:</p><ul><li>The user user01 logged into DC101</li><li>Logon type: <strong>Network</strong></li><li>Auth type: <strong>NTLM</strong></li><li>Same domain</li><li>Same user</li><li>Same authentication pattern</li></ul><p>This is exactly what:</p><ul><li>WMI execution</li><li>PSExec</li><li>NTLM relay</li><li>Remote service execution</li></ul><p>…looks like after the fact.</p><p>No console login.<br> No keyboard.<br> Just credentials moving.</p><p>From the graph, the inference was unavoidable:</p><blockquote><em>The attacker most likely pivoted from SRV01 into DC101 using NTLM credentials.</em></blockquote><p><em>Not proven by full packet capture — <br> but extremely strong confidence from identity correlations.</em></p><h3>The Final Attack Chain</h3><p>Reconstructed fully:</p><h3>1️⃣ NODE01</h3><p>PowerShell as administrator — breach point<br> ✅ Confirmed admin control</p><h3>2️⃣ SRV01</h3><p>PowerShell Remoting via WinRM (5985)<br> ✅ Confirmed lateral movement</p><h3>3️⃣ DC101</h3><p>NTLM network logons from the same user<br> 🟧 Highly likely pivot</p><h3>Why Graphs Made This Obvious</h3><p>Without the graph:</p><ul><li>Logs are separate</li><li>Systems look isolated</li><li>Events look “normal”</li></ul><p>With the graph:</p><ul><li>Relationships become visible</li><li>Movement becomes directional</li><li>Identity reuse becomes obvious</li></ul><p>This wasn’t log analysis.</p><p>This was <strong>path reasoning</strong>.</p><h3>What Would Confirm This 100%?</h3><p>To make this absolutely bulletproof I’d need:</p><ul><li>Event 4624 (logons)</li><li>Event 4688 (process creation)</li><li>Event 4672 (admin privileges)</li><li>PowerShell logs (4104)</li><li>WinRM logs</li><li>Scheduled task creation (7045)</li></ul><p>But even without them, the attack path was:</p><blockquote><em>Architecturally undeniable.</em></blockquote><h3>Final Verdict</h3><p>This was not a random incident.</p><p>It was a deliberate, multi-step lateral movement campaign using:</p><ul><li>PowerShell Remoting</li><li>NTLM credentials</li><li>Domain trust</li><li>Windows-native tools</li></ul><p>And the graph revealed it effortlessly.</p><h3>Closing Thought</h3><p>If you still think logs tell the full story…</p><p>You’re only seeing frames.</p><p>Graphs show the movie.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5bc9ce2830ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Detection of Windows Management Instrumentation (WMI)]]></title>
            <link>https://virenchoudhari.medium.com/detection-of-windows-management-instrumentation-wmi-e5349f1235b?source=rss-b8694a2687db------2</link>
            <guid isPermaLink="false">https://medium.com/p/e5349f1235b</guid>
            <dc:creator><![CDATA[Viren Choudhari]]></dc:creator>
            <pubDate>Sun, 18 Apr 2021 10:40:12 GMT</pubDate>
            <atom:updated>2021-04-18T10:40:12.368Z</atom:updated>
            <content:encoded><![CDATA[<p>WMI is a Microsoft implementation of Web-Based Enterprise Management (WBEM) in Windows OS</p><p>What is WBEM?</p><ul><li>It is an open standard called the Web-Based Enterprise Management (WBEM) which aims to exchange Comon Information Model (CIM) data between systems</li></ul><p>What is CIM?</p><ul><li>It is a data model standard which allows hardware, Operating Systems or applications for example to define classes that represent the different components they have. An example would be a class that represents information about hardisk</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e5349f1235b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>