<?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 MD Wakif Ibne Mizan on Medium]]></title>
        <description><![CDATA[Stories by MD Wakif Ibne Mizan on Medium]]></description>
        <link>https://medium.com/@wakif360?source=rss-6c9dd61c736f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*8pPq-UG1pWaKTaWr</url>
            <title>Stories by MD Wakif Ibne Mizan on Medium</title>
            <link>https://medium.com/@wakif360?source=rss-6c9dd61c736f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 21:57:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@wakif360/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[Technical failure in Japan’s banking system disrupts 128 financial institutions.]]></title>
            <link>https://medium.com/@wakif360/technical-failure-in-japans-banking-system-disrupts-128-financial-institutions-83f8dbae5b7e?source=rss-6c9dd61c736f------2</link>
            <guid isPermaLink="false">https://medium.com/p/83f8dbae5b7e</guid>
            <dc:creator><![CDATA[MD Wakif Ibne Mizan]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 08:02:33 GMT</pubDate>
            <atom:updated>2026-11-08T08:02:33.860Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kyuYQn2IBxm8tVjEYVyqJg.png" /></figure><p>It’s easy to imagine the horror of thousands of customers being unable to access their money because of a system failure. On November 6, 2026, 128 Shinkin banks throughout Japan (mainly in the western parts of the country) experienced service disruptions due to a failure in their shared banking infrastructure. Around 3430 branches were affected. Customers were unable to access services such as ATMs and counter services. Eventually, it was discovered that the issue had been caused by equipment failure at one of the centers of their shared banking infrastructure. While services were restored quickly, the situation highlighted a major flaw in our commonly connected banking infrastructure. If many different banks are using the same technology, issues can affect many people. So what happened, and how did one technical failure cause so many issues for banks?</p><p>At approximately 1:30 pm JST on the 6th of November 2026, there was an issue with the Shinkin Joint System that caused banking disruptions at 128 Shinkin banks located primarily in Western Japan. This includes branches in various regions such as Tokai, Hokuriku, Kinki, Chugoku, Shikoku, and Kyushu. According to reports from Japanese media outlets, around 3,430 branches were affected. This meant customers were unable to perform standard banking activities such as depositing or withdrawing cash, using over-the-counter banking services, and some online banking. The impacted banks include Okazaki Shinkin Bank (Aichi Prefecture), Osaka City Shinkin Bank, and Hiroshima Shinkin Bank. This was a significant disruption because whilst the banks were independent, they used the same banking system. This meant many banks were affected. However, not all Shinkin banks in Japan were impacted, as those in the East continued with normal operations. This localised banking issue caused a huge disruption for a large portion of Japan’s cooperative banking industry.</p><p>It is important to understand the common infrastructure of Japan’s Shinkin banking network to understand why many financial institutions were impacted. Shinkin banks all use a common network known as the Shinkin Joint System. This is an online network managed by the Shinkin Kyodo Centre. This helps banks within Shinkin perform vital functions such as account transactions, ATM services, and payment information sharing. On the 6th November, there was a failure within the network that was caused by a communications failure. As banking is reliant on constant communication between ATMs, branches and central systems, if there is a failure in a key part of the network, systems will not be able to share the information required to authorise and complete transactions. This can explain why many institutions were impacted around the same time. The Shinkin Kyodo Centre has stated that the issue was caused by a failure with the equipment and not a cyber attack. Reports have shown that it is unlikely the breach was due to unauthorised access. However, there is no public information on what caused the technical issues. The reports do not state what equipment failed, what the issue with it was, or if there were any backup communication paths in place. It has been reported that the institutions had recently upgraded their infrastructure in October. There is no proof of a connection between the upgrade and the outage.</p><p>The incident was resolved on the same day, but it took time to get everything back up and running. After the incident occurred around 1:30 pm on November 6, engineers at the Shinkin Kyodo Centre tried to get the network and banking services back up and running. Some of the ATMs and online banking services were brought back online, according to Japanese broadcaster TV Asahi. However, by around 7:05 pm Japan Standard Time, the Shinkin Kyodo Centre and National Association of Shinkin Banks said all services had been restored completely. This meant the incident had come to an end, after about 5-and-a-half hours. However, just because the network had been brought back online, it does not mean everything had been processed. Some of the transfers to other financial institutions had been delayed because of the incident and would be processed on the next working day (November 9). The system operator apologised for the disruption and said they would look into why the equipment failed to ensure it does not happen again. However, there had not been any specific details released about what measures would be taken to prevent it from happening again.</p><p>This raises the bigger picture question around the resilience of financial technology. This incident highlights risks around using centralised and shared infrastructure. Although using shared infrastructure increases efficiency, there is a risk of creating a single point of failure. This can cause issues for many organisations using the same service. Whilst this doesn’t mean the infrastructure is poorly designed, it does show the importance of regularly testing the resilience of the infrastructure to handle component failures.</p><p>There are a variety of technical lessons to learn from this incident. Firstly, redundancy shouldn’t just be implemented at a hardware level. Communication routes should be in place within the network, and redundancy should be in place to ensure essential functions within the network can still be maintained. Having redundant hardware won’t help if everything is relying on the same network, power source, and environment. Secondly, failover should be tested. Having a recovery solution in place is only useful if we can fail over reliably. Thirdly, monitoring and isolation should be considered as key functions. If we can monitor the hardware and identify unusual activity quickly, we can isolate the issue, which will reduce any downtime. Lastly, updates should be thoroughly validated. Following the incident in Japan, it was found that there could be a link between the hardware failure and an update to the ATM system in the western region in October. Whilst we are still investigating the exact reason for the failure, this shows the importance of managing changes to the system. We should ensure updates are tested in the correct configuration, have a rollback plan, and carefully monitor the system following the update.</p><p>This isn’t just an issue within banking. As organisations become more reliant on shared infrastructure such as cloud, payment networks, communications and centralised digital systems, the reliability of one technical provider can affect many organisations. This can also be an attractive target for malicious individuals. Although there was no evidence of a cyber attack in this case, this incident shouldn’t be taken as evidence of a breach. It should be seen as a reminder that availability, resilience and recovery are key elements of digital security. Whilst preventing unauthorised access is a responsibility, it’s also important to ensure essential services are reliable in unexpected technical situations.</p><p>The recent disruption caused by the Japanese banking industry serves as a great reminder that as we move forward with new technologies, we must ensure the reliability of our systems. As more and more financial services use interconnected systems, problems with the technology can cause issues in other areas. This incident shows that we should not just be focused on keeping our systems secure from malicious activity. We should also ensure that we have reliable systems, recovery processes, and manage any dependencies. Financial institutions should not just be looking at creating efficient systems but ensuring that they are reliable. In today’s banking world, we need to be able to create systems that are not only efficient but reliable in the event of any unexpected failure.</p><h4>Sources</h4><p>1. Nippon.com — Jiji Press: System Failure Disrupts ATMs at 128 Shinkin Banks in Japan<br><a href="https://www.nippon.com/en/articles/-/57039?utm_source=chatgpt.com">https://www.nippon.com/en/articles/-/57039</a></p><p>2. Nippon.com — Jiji Press: (Update) Glitch Briefly Disrupts ATMs at 128 Shinkin Banks in Japan<br><a href="https://www.nippon.com/en/articles/-/57057?utm_source=chatgpt.com">https://www.nippon.com/en/articles/-/57057</a></p><p>3. TBS NEWS DIG — Western Japan’s 128 Shinkin Banks Fully Restore Services After System Failure<br><a href="https://newsdig.tbs.co.jp/articles/withbloomberg/3068915?utm_source=chatgpt.com">https://newsdig.tbs.co.jp/articles/withbloomberg/3068915</a></p><p>4. Nikkansports — Kyodo News: System Failure Affects 128 Shinkin Banks in Western Japan<br><a href="https://www.nikkansports.com/general/news/202611060001790.html?utm_source=chatgpt.com">https://www.nikkansports.com/general/news/202611060001790.html</a></p><p>5. FNN Prime Online — 128 Shinkin Banks Fully Restore Transactions Following Equipment Failure<br><a href="https://www.fnn.jp/articles/-/1147822?utm_source=chatgpt.com">https://www.fnn.jp/articles/-/1147822</a></p><p>6. FinTech Journal — Communication Equipment Failure Disrupts 128 Shinkin Banks for Nearly Six Hours<br><a href="https://www.sbbit.jp/article/fj/188172?utm_source=chatgpt.com">https://www.sbbit.jp/article/fj/188172</a></p><p>7. Kobe Shimbun NEXT — Kyodo News: 128 Shinkin Banks in Western Japan Restore Services After Six-Hour Outage<br><a href="https://www.kobe-np.co.jp/news/kyodo/202611/0021008631.shtml?utm_source=chatgpt.com">https://www.kobe-np.co.jp/news/kyodo/202611/0021008631.shtml</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=83f8dbae5b7e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI vs AI: The Last 48 Hours of Cyber Conflict]]></title>
            <link>https://medium.com/@wakif360/ai-vs-ai-the-last-48-hours-of-cyber-conflict-3fbfa6d2d68e?source=rss-6c9dd61c736f------2</link>
            <guid isPermaLink="false">https://medium.com/p/3fbfa6d2d68e</guid>
            <dc:creator><![CDATA[MD Wakif Ibne Mizan]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 07:30:59 GMT</pubDate>
            <atom:updated>2026-10-24T08:26:29.463Z</atom:updated>
            <content:encoded><![CDATA[<p>The AI race has already begun. Both sides are using artificial intelligence (AI) to strengthen cybersecurity. However, the same technology that helps protect and defend can also be used to infiltrate and exploit. This makes the issue particularly pressing: there are too many incidents, vulnerabilities, and other events. A single point of attack can significantly damage the security of any IT infrastructure before specialists even have time to analyze and respond.</p><p>Thus, the situation requires more than tracking hacking attempts or discovering new vulnerabilities. Most importantly, it is necessary to consider what exactly is happening and how relevant it is to the AI arms race. Let’s take a look at the most critical events of the last 48 hours, one by one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1xQfOGBpscb8jna0H3Gc1w.png" /></figure><h4>Signals critical calling bug</h4><p>A bug in the signaling system of the messaging app Signal has raised privacy concerns about the ability of its users to terminate calls. According to The Register, an attacker could exploit the vulnerability and prevent calls from ending as intended. The problem reportedly exists in RingRTC, the open-source project for creating voice and video calls used by Signal. An attacker could send flood requests that would block any further actions, including hanging up a call, by overwhelming the internal event queue of the target device. According to the report, even though Signal employs end-to-end encryption, it cannot prevent the abuse of an unintended transmission during a failed call. However, the vulnerability is not deep enough to fully abuse a target’s microphone. The issue has been fixed in the latest version of RingRTC (2.72.0) as of writing this article.</p><h4>Zeabur Data Breach Exposes More Than 321,000 Accounts</h4><p>Cloud deployment platform Zeabur has announced a security breach that involved the theft of data from approximately <strong>321,810 accounts and over 3.2 million environment-variable</strong> records by an attacker. According to the company’s incident report, the breach occurred in October during unauthorized access to Zeabur’s shared infrastructure, platform database, and some of Google Cloud’s dedicated servers with the help of stolen credentials. The compromised data included API tokens, account information, project details, and environment variables that hold the passwords and other credentials to secret services. Zeabur’s report stated that the company had failed to perform a complete credential rotation and enhance the security of its services following an August breach. The platform announced that it had revoked the compromised secrets and recommended that users change the affected secrets if their projects were using the breached tokens.</p><p>Zeabur announced on October 22 that it would shut down its platform by the end of 2026 due to the challenges it faced with the breach and the security issues it had suffered. The report noted that the breach could see attackers gain access to all the customers’ projects and data on the cloud-deploying platform.</p><h4>LiteLLM Supply-Chain Attack Exposes Developer Credentials</h4><p>On October 21, 2026, attackers compromised the publish process of LiteLLM, a popular AI gateway, and distributed a backdoored version of the software, which could steal sensitive credentials. According to LiteLLM’s official incident report, version 1.107.0 was published to PyPI with compromised Docker images and was available for roughly two hours. Microsoft Security researchers determined that the malicious file could be executed when Python is launched, allowing it to steal user credentials from various developer tools, cloud platforms, browsers, and AI applications. Moreover, the attack surface was expanded, as it included a loader that could bypass protections and run other malicious payloads.</p><p>Furthermore, according to LiteLLM’s investigation, the attacker used stolen GitHub credentials to impersonate the official repository and push the distribution with malicious code. The compromised versions of the company’s software have been removed from PyPI, and users who have installed or run the affected release should consider their systems insecure, rebuild them, and rotate all compromised credentials. The report concluded that the incident showed how a compromised software update could damage systems unrelated to the affected application.</p><h4>US Banks Restrict AI Agents Amid Concerns Over Customer Accounts</h4><p>Major US banks, including JPMorgan Chase, Morgan Stanley, Bank of America and Wells Fargo, are restricting third-party AI agents’ access to customer accounts, <em>Axios and Reuters</em> reports. The move follows allegations that some AI agents, given customers’ login credentials, accessed banking platforms and appeared to take actions beyond users’ instructions. Some were said to have contacted a bank’s call centre without first consulting the account holder. There are doubts about the reliability of autonomous AI in dealing with sensitive financial information. JPMorgan CEO Jamie Dimon warned that unrestricted access by external agents is a risk to the customer and the bank. The institutions are restricting access and are considering safer alternatives, including read-only access. Notably, reports have not established that the incidents involved agents hacking into accounts or stealing funds.</p><h4>PHP2Shell: Critical Vulnerabilities Put Web Servers at Risk</h4><p>Two critical security flaws in PHP software, collectively named PHP2Shell, have been reported to impact the security of websites and web applications around the globe. According to BleepingComputer, the issues, tracked as CVE-2026–109618 and CVE-2026–109619, have been assigned a base score of 9.2 out of 10 and could enable remote code execution attacks to commence. Both flaws were reported to stem from memory-management errors in PHP that could be triggered during abnormal termination of requests. Exploitation of the vulnerabilities could enable threat actors to gain control over targeted server processes and steal sensitive data. The UK’s NHS England Digital advised that systems running vulnerable PHP versions should undergo patching as a matter of urgency. Emergency support came in the form of PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. There are yet to be any indications that all vulnerable web servers are under attack, despite reports of attempted exploitation. Administrators running vulnerable PHP versions are advised to update their installations and remain vigilant for signs of compromise.</p><h4>AI-Discovered PHP Bugs Force Government Websites Offline Worldwide</h4><p>Artificial intelligence has exposed two more significant vulnerabilities that were concealed in the PHP programming language, which runs about 70–75 percent of the identified websites. Reuters reports that the problems prompted governments and organizations in Europe and Japan to take down some websites and online services for the time being. The flaws tracked down as <strong>CVE-2026–109618 and CVE-2026–109619</strong> could enable attackers to execute arbitrary code on the server where the targeted websites are hosted, without authentication. The researchers at <em>Calif and Wiz</em> cybersecurity firms spotted the issues with the help of AI-driven tools that were used to scan for weaknesses that remained undiscovered for a considerable period of time. Subsequently, authorities urged firms to either stop using vulnerable software or cut off servers from the public internet in order to minimize risks. According to the Guardian, millions of attacks have been launched at the vulnerable systems as organizations scrambled to mitigate the damage. Among the affected ones are the Dutch government domains, Finland’s state-wide IT platforms, and several ministry sites in Japan</p><h4>Node.js Security Flaw Discovered with AI Raises Remote-Code-Execution Risks</h4><p>On October 21, the Node.js project released security updates for 10 vulnerabilities in the 22.x, 24.x, and 26.x versions. The most serious was CVE-2026–111497, an elevation of privilege flaw in the Brotli decompression part, which could lead to memory corruption, information leaks, or even remote code execution with specially crafted compressed data. The Node.js security team said <em>Khai Tran and Sylvie</em> Mayer from <em>OpenAI Codex Security</em> found the vulnerability. This shows how large language models can help find issues in popular software. However, these discoveries don’t always lead to immediate attacks — a proof-of-concept only worked on a weakened test build, not a normal Node.js binary. Security fixes are now available in versions 22.23.4, 24.22.1, and 26.12.1.</p><h4>GitHub Fixes Critical Vulnerability in Its Git Push Pipeline</h4><p>GitHub also dealt with a major security issue after Wiz researchers found a vulnerability in its Git push processing pipeline. This flaw could let users with push access run any commands on the server handling the operation, putting repositories and developer infrastructure at risk. GitHub’s security report said the problem was due to not properly sanitizing certain Git operations. GitHub quickly fixed the issue on GitHub.com and found no evidence it was exploited. Security updates were also released for supported GitHub Enterprise Server versions.</p><h4>PingCAP Breach: Attackers Claim 10.6 TB of Data Stolen</h4><p>PingCAP, which makes the TiDB distributed SQL database, reported a security breach where an unknown attacker used stolen credentials to get into the system. The company’s October 2026 report says the breach happened between September 9 and October 7, during which the attacker copied information and stole source code. The stolen data includes customer accounts, project info, SQL editor text, errors, business records, and encrypted API keys. PingCAP says there’s no evidence that customer data in the TiDB cloud was breached. The incident got more attention after a group called FulcrumSec claimed to have taken 10.6 EB of data in 4.9 million files, but PingCAP denied these claims and said the attacker did not use plain-text credentials.</p><h4>SBI Data Leak Exposes Personal Information Linked to 3,897 Customer Records</h4><p>The State Bank of India (SBI) says there may have been a data leak after finding customer information on the dark web. The bank found a file with customer data circulating on dark web forums. Reuters reports that out of 375,000 entries checked, 3,897 had personal information like names, mobile numbers, birth dates, emails, addresses, account numbers, and loan details. SBI hasn’t confirmed the data or its source yet. The bank has started an investigation and hired an external cybersecurity firm. On October 23, SBI notified affected customers by name and assured them their money is safe. The bank is warning all customers to watch out for phishing scams using the stolen data.</p><h4>Microsoft Reports China-Linked Hackers Using AI Agents in Global Cyberattacks</h4><p>On October 21, 2026, Microsoft said that a China-linked group called Storm-0940 used ARTEX, an AI penetration-testing tool, to automate cyberattacks on hundreds of organizations worldwide. Targets included government, finance, energy, manufacturing, and tech sectors. Microsoft Threat Intelligence reported that ARTEX automated tasks like reconnaissance, password-spraying, and credential abuse, and in some cases, attackers got privileged access to cloud infrastructure in under three hours. Microsoft connected this activity to the ConnectWise Cloud Backup supply-chain breach, where attackers used legitimate app access to reach downstream customers. While most attacks failed, Microsoft saw successful breaches and persistence in several organizations. Microsoft warned that AI automation can greatly shorten the time defenders have to spot and stop intrusions. The report showed that agentic AI can do more than suggest attacks — it can coordinate several steps of an intrusion.</p><h4>AI Is Changing Reverse Engineering and the Cybersecurity Race</h4><p>Reverse engineering has long been the domain of the specialist with the task of analysing binaries, working out what the program does, looking at firmware for weaknesses that you simply can’t see in the source code, and so on and so forth. AI is now beginning to assist with some of the work and is in fact changing the nature of the discovery process such that we’re now required to respond much more quickly.</p><p>In Oct 2026, Microsoft announced that its FORGE Lab had contributed to the discovery of 140 vulnerabilities in Windows and 155 internal reports, which were validated for 23 open source projects. Microsoft feels that it’s not just about finding the issues; it’s about validating the findings and then turning that into a security fix, as the volume of issues being found automatically is such that the process has to keep pace. Google has also reported the use of its Gemini 4 Argon model to find and fix bugs in its own and open-source software. One of the bugs found by Gemini 4 Argon was a critical vulnerability in hospital software which could potentially result in the exposure of hundreds of thousands of patient records. The use of large language models to research software weaknesses is a two-edged sword in that it will undoubtedly make the programmer safer in terms of being able to analyse and learn the code and thus find the hidden flaws. It can, of course, also be used by the bad guys to rapidly research the weaknesses in software.</p><p>Another example of large-scale AI being used in security research is OpenAI’s contribution to the discovery of a problem in the Brotli compression libraries, which impacted Node.js and a number of other software packages. In this instance both the issue and the fix were found by having the model run in a constrained environment (again, fairly standard for research). Faster security research will of course allow us to find more issues, but it will also raise the question of how quickly we can respond and how quickly we can deploy a patch to stop an attacker from exploiting the vulnerability once it’s been discovered. You have to validate the issue, you have to work out the impact of the vulnerability, you have to find a fix, and you have to actually implement that fix. All of that takes time even with AI-assisted development. The fundamental question of whether the defenders can keep up with the attackers is going to be the central one in the race to find and fix software vulnerabilities.</p><h4>Conclusion</h4><p>AI is providing a significant boost to reverse engineering and cybersecurity. It allows for much faster analysis of software and the discovery of vulnerabilities, which in turn enables security researchers to find the flaws that would otherwise have taken a great deal of time to find. The same advantage that makes life so much easier for security researchers can be exploited by attackers to gain an understanding of a system they are unfamiliar with and to discover zero-days. The examples in the article illustrate that the real issue is not simply about working out what can be broken, it’s about how quickly those issues can be validated, reported, and patched before the bad guys can exploit them.</p><p>Statements from the big tech companies are unequivocal in their assertion that the new opportunities for security research that AI presents are very promising. The reality is that we do still have a requirement to improve our response processes. Even when AI can identify new vulnerabilities at a very rapid rate, we still need to go through the process of verification, responsible disclosure, and remediation. So the real question isn’t whether researchers can use AI to find new vulnerabilities more quickly; it’s whether the rate at which they can find them can keep up with the rate at which they can then respond to and fix each one.</p><h3>Sources</h3><p>1. Microsoft Security Blog: 3 Lessons from Frontier AI Vulnerability Research<br><a href="https://www.microsoft.com/en-us/security/blog/2026/10/07/3-lessons-from-frontier-ai-vulnerability-research/?utm_source=chatgpt.com">https://www.microsoft.com/en-us/security/blog/2026/10/07/3-lessons-from-frontier-ai-vulnerability-research/</a></p><p>2. Google: Developing Gemini 4 Argon Responsibly and Safely<br><a href="https://blog.google/innovation-and-ai/technology/safety-security/gemini-4-model-safety-security/?utm_source=chatgpt.com">https://blog.google/innovation-and-ai/technology/safety-security/gemini-4-model-safety-security/</a></p><p>3. Reuters: Reporting on OpenAI’s frontier AI security research and software vulnerability discovery<br><a href="https://www.reuters.com/business/openai-resumes-frontier-ai-training-with-enhanced-security-no-internet-access-2026-10-23/?utm_source=chatgpt.com">https://www.reuters.com/business/openai-resumes-frontier-ai-training-with-enhanced-security-no-internet-access-2026-10-23/</a></p><p>4. InCyber: AI Accelerates Reverse Engineering of Connected Devices<br><a href="https://incyber.org/en/article/ai-accelerates-reverse-engineering-of-connected-devices/?utm_source=chatgpt.com">https://incyber.org/en/article/ai-accelerates-reverse-engineering-of-connected-devices/</a></p><p>5. JPCERT/CC: Report on AI-assisted cyberattacks, API attacks, and mobile-application analysis<br><a href="https://blogs.jpcert.or.jp/en/2026/10/web_api_attacks.html?utm_source=chatgpt.com">https://blogs.jpcert.or.jp/en/2026/10/web_api_attacks.html</a></p><p>6. TechRadar: Reporting on AI-assisted reverse engineering of consumer PC accessories<br><a href="https://www.techradar.com/ai-platforms-assistants/claude/amazon-security-engineer-hacks-pc-accessories-with-claude-opus-to-make-them-work-better-asus-insta360-and-elgato-products-reverse-engineered-in-hours-for-better-control-but-engineer-admits-this-also-scares-me?utm_source=chatgpt.com">https://www.techradar.com/ai-platforms-assistants/claude/amazon-security-engineer-hacks-pc-accessories-with-claude-opus-to-make-them-work-better-asus-insta360-and-elgato-products-reverse-engineered-in-hours-for-better-control-but-engineer-admits-this-also-scares-me</a></p><p>7. Node.js Project: October 2026 Security Releases<br><a href="https://nodejs.org/en/blog/vulnerability/october-2026-security-releases?utm_source=chatgpt.com">https://nodejs.org/en/blog/vulnerability/october-2026-security-releases</a></p><p>8. The Hacker News: Node.js Brotli Decoder Bug Could Allow Remote Code Execution in All Supported Versions<br><a href="https://thehackernews.com/2026/10/nodejs-brotli-decoder-bug-could-allow.html?utm_source=chatgpt.com">https://thehackernews.com/2026/10/nodejs-brotli-decoder-bug-could-allow.html</a></p><p>9. Singapore Cyber Security Agency: Critical Vulnerability in Node.js<br><a href="https://www.csa.gov.sg/alerts-and-advisories/alerts/al-2026-154/?utm_source=chatgpt.com">https://www.csa.gov.sg/alerts-and-advisories/alerts/al-2026-154/</a></p><p>10. GitHub Blog: Securing the Git Push Pipeline: Responding to a Critical Remote Code Execution Vulnerability<br><a href="https://github.blog/security/securing-the-git-push-pipeline-responding-to-a-critical-remote-code-execution-vulnerability/?utm_source=chatgpt.com">https://github.blog/security/securing-the-git-push-pipeline-responding-to-a-critical-remote-code-execution-vulnerability/</a></p><p>11. The Verge: Reporting on GitHub’s response to a critical remote-code-execution vulnerability<br><a href="https://www.theverge.com/news/920295/github-remote-code-execution-vulnerability-fix?utm_source=chatgpt.com">https://www.theverge.com/news/920295/github-remote-code-execution-vulnerability-fix</a></p><p>12. Reuters: Government Websites Go Dark After AI-Discovered Flaw Sparks Security Scramble<br><a href="https://www.reuters.com/world/asia-pacific/eu-agency-warns-websites-may-need-be-taken-offline-php-prepares-patches-2026-10-21/?utm_source=chatgpt.com">https://www.reuters.com/world/asia-pacific/eu-agency-warns-websites-may-need-be-taken-offline-php-prepares-patches-2026-10-21/</a></p><p>13. The Guardian: Reporting on attempts to exploit the PHP vulnerabilities<br><a href="https://www.theguardian.com/technology/2026/oct/22/websites-across-world-hit-by-millions-of-attempts-to-exploit-php-code-flaw?utm_source=chatgpt.com">https://www.theguardian.com/technology/2026/oct/22/websites-across-world-hit-by-millions-of-attempts-to-exploit-php-code-flaw</a></p><p>14. ENISA:Joint Statement on Critical PHP Vulnerabilities<br><a href="https://www.enisa.europa.eu/news/joint-statement-on-critical-php-vulnerabilities?utm_source=chatgpt.com">https://www.enisa.europa.eu/news/joint-statement-on-critical-php-vulnerabilities</a></p><p>15. BleepingComputer: PHP Ships Emergency Patches for Critical PHP2Shell Flaws<br><a href="https://www.bleepingcomputer.com/news/security/php-ships-emergency-patches-for-critical-php2shell-flaws/?utm_source=chatgpt.com">https://www.bleepingcomputer.com/news/security/php-ships-emergency-patches-for-critical-php2shell-flaws/</a></p><p>16. Euronews: PHP Security Flaw: Why Websites Across Europe Went Dark and What It Means for You<br><a href="https://www.euronews.com/2026/10/22/php-security-flaw-why-websites-across-europe-went-dark-and-what-it-means-for-you?utm_source=chatgpt.com">https://www.euronews.com/2026/10/22/php-security-flaw-why-websites-across-europe-went-dark-and-what-it-means-for-you</a></p><p>17. Microsoft Security Blog: AI-Driven, Large-Scale Nation-State Campaign Using ARTEX Enables Supply Chain Compromise<br><a href="https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/?utm_source=chatgpt.com">https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/</a></p><p>18. Reuters: Reporting on China-linked hackers, ConnectWise, and attacks against hundreds of organisations<br><a href="https://www.reuters.com/technology/chinese-linked-hackers-compromise-connectwise-target-hundreds-global-attack-2026-10-21/?utm_source=chatgpt.com">https://www.reuters.com/technology/chinese-linked-hackers-compromise-connectwise-target-hundreds-global-attack-2026-10-21/</a></p><p>19. PingCAP: A Note from Our CEO on Our Security Incident<br><a href="https://www.pingcap.com/ko/trust-hub/security/ceo-note-october-2026/?utm_source=chatgpt.com">https://www.pingcap.com/ko/trust-hub/security/ceo-note-october-2026/</a></p><p>20. NavsPlace: TiDB Cloud Breach: What PingCAP Confirmed and What the Attacker Claims<br><a href="https://www.navsplace.net/tidb-cloud-breach-pingcap-20261022/?utm_source=chatgpt.com">https://www.navsplace.net/tidb-cloud-breach-pingcap-20261022/</a></p><p>21. AIIDeList: PingCAP Data Breach 2026: TiDB Cloud Impact and Response<br><a href="https://aiidelist.com/blog/pingcap-tidb-cloud-data-breach-2026?utm_source=chatgpt.com">https://aiidelist.com/blog/pingcap-tidb-cloud-data-breach-2026</a></p><p>22. Reuters: Reporting on the suspected exposure of SBI customer records<br><a href="https://www.reuters.com/legal/litigation/state-bank-india-reports-cyberattack-involving-customer-records-2026-10-21/?utm_source=chatgpt.com">https://www.reuters.com/legal/litigation/state-bank-india-reports-cyberattack-involving-customer-records-2026-10-21/</a></p><p>23. The Economic Times: SBI Data Leak Scare: Bank Contacts Affected Customers. Is Your Money Safe?<br><a href="https://economictimes.indiatimes.com/industry/banking/finance/banking/sbi-data-leak-scare-bank-contacts-affected-customers-is-your-money-safe/articleshow/135139356.cms?from=mdr&amp;utm_source=chatgpt.com">https://economictimes.indiatimes.com/industry/banking/finance/banking/sbi-data-leak-scare-bank-contacts-affected-customers-is-your-money-safe/articleshow/135139356.cms?from=mdr</a></p><p>24. LiteLLM: October 21, 2026 Security Incident<br><a href="https://docs.litellm.ai/blog/security-incident-oct-21-2026?utm_source=chatgpt.com">https://docs.litellm.ai/blog/security-incident-oct-21-2026</a></p><p>25. Microsoft Security Blog: Reporting on the LiteLLM 1.107.0 supply-chain compromise<br><a href="https://www.microsoft.com/en-us/security/blog/2026/10/23/inside-the-litellm-1-107-0-supply-chain-compromise-credential-theft-through-python-startup-hooks/?utm_source=chatgpt.com">https://www.microsoft.com/en-us/security/blog/2026/10/23/inside-the-litellm-1-107-0-supply-chain-compromise-credential-theft-through-python-startup-hooks/</a></p><p>26. The Register: Reporting on the LiteLLM supply-chain attack and credential theft<br><a href="https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/?utm_source=chatgpt.com">https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/</a></p><p>27. The Register: Signal Bug Could Keep Your Mic — and Camera — On After Hangup<br><a href="https://www.theregister.com/security/2026/10/21/signal-bug-could-keep-your-mic-and-camera-on-after-hangup/?utm_source=chatgpt.com">https://www.theregister.com/security/2026/10/21/signal-bug-could-keep-your-mic-and-camera-on-after-hangup/</a></p><p>28. Zeabur: October 2026 Security Incident<br><a href="https://zeabur.com/blogs/october-2026-security-incident?utm_source=chatgpt.com">https://zeabur.com/blogs/october-2026-security-incident</a></p><p>29. Pasquale Pillitteri: Zeabur Shuts Down After Two Breaches, 3.2 Million Records Read<br><a href="https://pasqualepillitteri.it/en/news/26411/zeabur-shuts-down-two-breaches-3-million-records?utm_source=chatgpt.com">https://pasqualepillitteri.it/en/news/26411/zeabur-shuts-down-two-breaches-3-million-records</a></p><p>30. Axios: Unvetted AI Agents Are Accessing U.S. Bank Accounts<br><a href="https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access?utm_source=chatgpt.com">https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access</a></p><p>31. Reuters: Reporting on US banks restricting third-party AI agents’ access to customer accounts<br><a href="https://www.reuters.com/legal/government/us-banks-detect-unauthorized-ai-agents-customer-accounts-axios-reports-2026-10-22/?utm_source=chatgpt.com">https://www.reuters.com/legal/government/us-banks-detect-unauthorized-ai-agents-customer-accounts-axios-reports-2026-10-22/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3fbfa6d2d68e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Information vs. Understanding: Rethinking AI Through Claude Shannon’s Information Theory]]></title>
            <link>https://medium.com/@wakif360/information-vs-understanding-rethinking-ai-through-claude-shannons-information-theory-9582251770f6?source=rss-6c9dd61c736f------2</link>
            <guid isPermaLink="false">https://medium.com/p/9582251770f6</guid>
            <category><![CDATA[information-theory]]></category>
            <dc:creator><![CDATA[MD Wakif Ibne Mizan]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 17:39:13 GMT</pubDate>
            <atom:updated>2026-10-23T17:39:13.098Z</atom:updated>
            <content:encoded><![CDATA[<p>If a machine gives you an answer, does that mean it understands the question?</p><p>This is the kind of question that comes from the growth of artificial intelligence (AI). The systems we call AI today can solve problems, explain many ideas, write essays that read like a human, and grasp the meaning of conversation. Because of this, it is hard to tell the difference between processing and true understanding.</p><p>A deeper problem lies behind these worries. Does processing data equal having an understanding of data when we talk about AI? To answer that, let us look back at the beginnings of information theory. Claude Shannon developed this theory to set rules for how to encode data. It does not deal with understanding at all. Still, the theory can be used to analyze and evaluate AI processing power and its limitations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OTh2s8xwWvg5cNPv.png" /></figure><p>In 1948, <strong>Claude Shannon</strong> published a paper titled “<em>A Mathematical Theory of Communication”</em>, which laid the foundation for modern <strong>Information Theory</strong>. The goal of his research was to define how information could be quantified and encoded in a way that allows for reliable transmission in the context of communication systems. Shannon was primarily interested in understanding how messages that carry information from a sender to a receiver could be encoded to reduce errors during transmission. Shannon’s paper focused on how information relates to the degree of uncertainty that it resolves. For instance, if a person already knows what another person is going to say, then the information received will have little bearing on their existing knowledge. In contrast, one’s degree of uncertainty is reduced when one receives unexpected information. In mathematical terms, the amount of information present is equal to the reduction in uncertainty, which Shannon called entropy. As Shannon later demonstrated, this approach could be used in practice to encode messages received from communication channels with encoding algorithms that minimize the possibility of errors during transmission. Information Theory, as defined by Shannon, focuses on information sources, their encodings, specific communication channels, and how information is ultimately received by a destination. Crucially, it must also be noted that Shannon’s definition of information does not consider the meaning of a received message. This is an important distinction, as Shannon’s communication model could, in principle, transmit nonsensical information that has no verifiable correspondence in the real world. In other words, Shannon’s Information Theory focuses on information entropy, not the information content itself. Now that we have seen how information is defined in the context of Information Theory, we can look at how similar concepts are being used in modern research, namely in the field of artificial intelligence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/957/0*wLLQgKaMxslgmQ5O.jpeg" /></figure><p>When asked a question by a user, an AI model can give a response that sounds like something humans might say. The model, for instance, can provide an explanation or clarification; it can connect ideas and relate facts of different natures, or it can produce an extensive step-by-step solution to the problem. But how can a machine perform such acts?</p><p>A language model, as it is called, is often compared to a statistical machine that learns a set of patterns from a large amount of data. This machine then uses these patterns to predict an outcome that is plausible, or probable, given the input it has received. The input is interpreted as combinations of words and word parts, called tokens. By repeating the process of learning and prediction, one can obtain a full response to a query posed to the model, with each token predicted step-by-step. But language models (LMs) are more interesting than meets the eye. Not only can they incorporate information from numerous sources in their predictions, but the statistical patterns they learn can also represent complex sets of information. This can enable them to answer questions, describe and explain problems in different ways, and perform calculations, even if the model has not seen a specific question or calculation before. The field of Information Theory, of which a large part deals with statistical patterns in text, can be of use when trying to understand what LMs do. The amount of information two texts (or tokens) share can be measured in terms of probability. This means that the patterns learned and stored by LMs can be useful in more ways than merely producing a realistic-sounding text. In practice, this can mean generating valid-sounding explanations for calculations performed by the model, or providing descriptive text for images, videos, and other media uploaded alongside a request.</p><p>But can such a system, however large and complex, actually understand the text it produces? To find out, one would have to investigate further than the responses given by the model. It would be necessary to analyze the patterns in the tokens that make up the generated text. Say someone says, “The bank is crowded.” Now, without any other context, do we know if the person means a financial institution or the shore of a river? The words are clear, the grammar is correct, and the sentence is easy enough to process and understand. However, understanding what is being said is a different matter.</p><p>This is the difference between processing information and interpreting. It’s one thing for a system to be able to identify patterns and connections, and to process and reply to information based on those patterns. But can it actually interpret the information? Statistical patterns in information are valuable resources, but their existence is not proof of anything other than themselves. As useful as patterns are in general, and as useful they can make answers, an answer that sounds right is not always an answer that is right. A confidence in pattern can easily become a misplaced certainty. Now, I’m not saying that pattern doesn’t prove something. I’m saying that pattern-based answers aren’t necessarily proof of comprehension. That’s not to say comprehension isn’t happening; it’s just to say that comprehension is not as simple as pattern-matching. The ability to answer questions fluently, using patterns in language, is not enough to tell us anything about whether the computer actually understands anything at all.The question of how to distinguish between mere pattern-following and an actual ability to comprehend and interpret meaning is where Shannon’s Information Theory comes in handy, providing us with a baseline measurement of information and uncertainty. Now, while calculating information content and uncertainty in a message is an excellent start, this is still only the beginning of the discussion about comprehension.</p><p>It would be easy to assume that AI does not understand anything at all, because it only knows statistics. But that would be as wrong as believing that only because it speaks fluently, it understands everything. Modern AI can do much more than repeat the same phrases. It can combine different contexts, create associations between them, react differently to the same phrase in different contexts, solve new problems, and apply analogies. But then, if we are talking about understanding, can we say that this is not understanding? Or, in other words, if a system has skill in applying information, selecting the right phrases, and solving new problems using old information, is it not understanding?</p><p>It depends on how we define understanding. Based on existing capabilities, we can confidently conclude that the AI can understand the context in which it was trained. But if we are talking about comprehension and the ability to grasp associated and remote connections, the question is more complicated. The problem is that neither of these two options can be taken as an argument against artificial intelligence understanding. While the first case applies only to speech, where AI has skills similar to humans, but not necessarily understanding, the second case is more complex. It cannot be argued that AI does not understand because its internal mechanisms are different from humans. However, it can be discussed whether AI demonstrates signs of comprehension comparable to human ones. Most likely, the discussion goes too far in both arguments. It would be much more productive to try to answer the question of what signs of understanding can be called understanding. Then, instead of denying or affirming it, you can look at the evidence that AI can demonstrate.</p><p>If Shannon’s Information Theory fails to explain our understanding, will it mean that mathematics does not have anything more to say? Not necessarily. It gives us a good basis to study information. But perhaps we need to look into information beyond its quantity, to see if there is something in it that would enable understanding. What we may seek is semantic information, which considers not only the statistics of information but its meaning and grounding, or the ability of an AI to associate words with objects, events, and states of affairs outside of language. We might need to look past just the amount of information and test how well AI understands what it says and how it uses that knowledge. We should see if it can interpret similar ideas, spot ambiguity, and use what it knows in different ways. These tests won’t prove AI understands like humans, but they go deeper than basic statistical analysis. The problem is that every idea of understanding comes with its own criteria for judging if AI has reached that level. Researchers can debate which definition is best, but each method needs its own evidence, and that evidence might not be enough to prove understanding exists. Maybe the goal isn’t to find a single formula for understanding. Instead, we should create criteria to judge what an AI can do, what it has shown it can do, and what it hasn’t done yet.</p><p>A machine can analyze data, learn patterns, and give a response that seems reasonable. But does that really tell us much about its ability to understand? Maybe not. Shannon’s information theory teaches us about information, uncertainty, and communication. But when it comes to today’s AI and what’s next, the questions themselves might be more interesting. Personally, I don’t think we should dismiss research or claim that AI’s abilities mean it truly understands. Instead, we should look at what it can do, how well it does it, and then draw our conclusions.</p><p>Ultimately, the debate concerns not only the capabilities of machines but also the broader conception of understanding. Although information theory may not soon resolve the question of what constitutes understanding, it continues to inform our understanding of the potential for machine comprehension.</p><h4>References</h4><p>1) C. E. Shannon, “A Mathematical Theory of Communication,” <em>The Bell System Technical Journal</em>, vol. 27, no. 3, pp. 379–423, July 1948, and no. 4, pp. 623–656, October 1948. Link: <a href="https://doi.org/10.1002/j.1538-7305.1948.tb01338.x">https://doi.org/10.1002/j.1538-7305.1948.tb01338.x</a></p><p>2) M. Mitchell and D. C. Krakauer, “The Debate Over Understanding in AI’s Large Language Models,” <em>Proceedings of the National Academy of Sciences</em>, vol. 120, no. 13, e2215907120, 2023. Link: <a href="https://doi.org/10.1073/pnas.2215907120">https://doi.org/10.1073/pnas.2215907120</a></p><p>3) “Information Rate of Meaningful Communication,” <em>Proceedings of the National Academy of Sciences</em>, vol. 122, no. 25, e2502353122, 2025. Link: <a href="https://doi.org/10.1073/pnas.2502353122">https://doi.org/10.1073/pnas.2502353122</a></p><p>4) H. Chen, X. Yang, J. Zhu, and W. Wang, “Quantifying Semantic Emergence in Language Models,” <em>Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics</em>, 2025. Link: <a href="https://aclanthology.org/2025.acl-long.588/">https://aclanthology.org/2025.acl-long.588/</a></p><p>5) D. A. Haslett and Z. G. Cai, “How Much Semantic Information is Available in Large Language Model Tokens?” <em>Transactions of the Association for Computational Linguistics</em>, vol. 13, pp. 408–423, 2025. Link: <a href="https://doi.org/10.1162/tacl_a_00747">https://doi.org/10.1162/tacl_a_00747</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9582251770f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Potential Limitations of AI Guardrails Through the Lens of W.]]></title>
            <link>https://medium.com/@wakif360/the-potential-limitations-of-ai-guardrails-through-the-lens-of-w-d891b0b7e823?source=rss-6c9dd61c736f------2</link>
            <guid isPermaLink="false">https://medium.com/p/d891b0b7e823</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cybernetics]]></category>
            <dc:creator><![CDATA[MD Wakif Ibne Mizan]]></dc:creator>
            <pubDate>Sat, 17 Oct 2026 07:56:02 GMT</pubDate>
            <atom:updated>2026-10-17T07:56:02.667Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Potential Limitations of AI Guardrails Through the Lens of W. Ross Ashby’s Law of Requisite Variety</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/739/1*gpggth2DeF-r7Q_Kdh90MA.jpeg" /></figure><p>What if the number of ways a system can be challenged exceeds the rules provided to protect it? In Artificial Intelligence (AI), the safety boundary that protects the system is known as AI guardrails. Despite these strict restrictions, many prompts and creative attempts often challenge guardrails. This raises some questions about guardrails: do they have the capability to handle all possible guardrail-breaking inputs? Decades ago, the Law of Requisite Variety, proposed by British cybernetician W. Ross Ashby, concerns disturbance variety and the regulator’s capability. In light of that law, we can understand and validate the limitations to some extent.</p><h4>Understanding the law</h4><p>William Ross Ashby (1903–1972) was a British psychiatrist and a pioneer of cybernetics. Cybernetics is the study of control and communication in complex systems. In his 1956 book, An Introduction to Cybernetics, Ashby formally introduced the Law of Requisite Variety, which describes the relationship between disturbances and a system’s ability to regulate them. [1]</p><p>In Simpler terms, the law states that a regulator requires sufficient variety of responses in order to manage the variety of disturbances it faces. Let’s imagine there is a switch for the room temperature controller that can only be turned on or off, but can’t switch between temperatures, so the variety of the input options is less than the possible inputs. Its environmental regulation ability is insufficient.</p><p>Ashby expressed this in a line: “Only variety can destroy variety.” [1] Here, the variety refers to the range of possible encounters by a system. A unique response is not mandatory for every encounter. Rather, the system must have the capabilities to face all of them to handle relevant differences.</p><h4>Understanding AI guardrails</h4><p>AI guardrails refer to the specific, determined boundaries for an AI model. They restrict AI models from harmful or unauthorized behaviour, focusing on appropriate responses. They can include input filters, output checks, access restrictions, and rules governing how an AI system uses external tools. [2]</p><p>For instance, a chatbot will be configured not to reveal sensitive information, and a healthcare AI assistant may be restricted from providing dangerous medical advice. Similarly, a company’s AI coding assistant may restrict certain operations or require additional checks before generating or executing risky code. [2] [3]</p><p>Guardrails can be effective in many different stages: before input reaches the AI model, while processing the request, after generating responses, etc. In AI agents, many files and APIs may also get restricted. These measures help manage risks, but their effectiveness depends on how they are designed, implemented, and tested. [2] [3]</p><h4>Relation between AI guardrail limitations and the Law of Requisite Variety</h4><p>W.Ross Ashby’s Law of Requisite Variety points out some ways to examine the chances of AI guardrails struggling against attempts to bypass them. According to the recommendation of the law, we can say that the regulator must contain sufficient ways to defend against all possible disturbances. [1]</p><p>Now, applying this law to AI guardrails, we can assume prompt injections and security layer bypasses as some possible disturbances. An attacker can embed a harmful command disguised as normal text or enquiry. If the AI guardrail model depends on a specific set of possibilities and predictable patterns only, it may be bypassed easily by an unfamiliar type of attack. [2] [3]</p><p>But that doesn’t make it necessary for the guardrails to have defensive infrastructure for every prompt. Meanwhile, the possible number of inputs in AI is practically limitless. A user can create seemingly unlimited combinations of words, numbers, attachments, and content that may look seemingly harmless. But in comparison, the possible ways to defend them are finite. This forms a difference between the number of varieties of attacks and the variety of defences.</p><p>Research on prompt injection has highlighted the challenges of designing and evaluating defences against diverse attacks.[4]. From the perspective of the Law of Requisite Variety, the question is whether an AI system can defend against all possible guardrail bypass attempts.</p><h4>Pointing out the AI guardrail limitations</h4><p>From the lens of W.Ross Ashby’s Law of Requisite Variety, several guardrail limitations in AI can be distinguished. The limitations don’t refer to ineffectiveness; they point out evolving challenges in defending against bypass attacks.</p><p>1) Unpredictable input variation: A user can provide similar intentions using combinations of many factors, indirect wording, language, etc. A safeguard protocol may defend against nearly all possible attacks, but can struggle with variation not learned. [2] [4]</p><p>2) Context-dependent interpretation: A similar phrase can affect differently in different scenarios. Guardrails should identify harmful requests without blocking necessary conversations. The identification becomes tough when the harmful command is hidden within a large textual context.</p><p>3) False positives and false negatives: Guardrails may distinguish normal requests as harmful, and harmful requests as normal. These errors reflect the difficulty of distinguishing acceptable from unacceptable content consistently, especially when intent depends on context [2].</p><p>4) Limited protections against connected tools: When an AI model is connected to multiple applications or APIs, it can be attacked by malicious conduct from the source. So, just focusing on the user’s input is not enough. [2] [3]</p><p>5) Evolving attack techniques: The AI guardrail systems must be kept under regular audits to make them defensive for all new types of attacks possible. The defensive system should be under consistent development through assessments. [4]</p><p>These possible limitations arise from reviewing Ashby’s Law of Requisite Variety.</p><h4>What Direction Does the Law Offer Rather Than a Complete Solution?</h4><p>The study and understanding of W.Ross Ashby’s Law of Requisite Variety encourages developers to work with the process of identifying the possible threats the AI model might encounter and create defensive measures for them. It is recommended to stop relying on fixed rules and broaden the defensive spectrum. They can use layered protection rather than a single protection mechanism.</p><p>Another direction is continuous development and evaluation. The variety of new attacking patterns will keep emerging, so the guardrails must be developed over time at regular intervals too. The National Institute of Standards and Technology (NIST) similarly emphasizes ongoing risk management, testing, evaluation, and monitoring throughout the AI lifecycle [3].</p><p>However, the Law of Ashby doesn’t provide any specific framework concept. It provides the idea of having the attack and defense variety equal in quantity to provide a rigid guardrail system in AI. The complete idea is conceptual but effective if kept in mind during system development.</p><h4>Conclusion</h4><p>The Law of Requisite Variety by W. Ross Ashby provides a broader perspective in terms of fighting against the emerging security threats in AI systems. Not directly, but conceptually. With more diverse prompting styles, security must be enhanced at an equal pace of emergence. However, I believe that Ashby’s law doesn’t prove that current AI guardrail systems are destined to fail. The current systems are clearly strong enough to fight most of the possible threats. But the concept rather explores the further emerging possibilities, the characteristics of the pace of upgrades, and the differences in the variety of risks and defenses.</p><p>The perspective raises many new thoughts and questions and builds an interesting conceptual connection between cybernetics and AI security. This perspective also aligns with broader AI security research, which classifies adversarial attacks and mitigation strategies to help researchers understand and address different threat types [5]. Finally, I believe the law of Ashby is not about predicting AI guardrail failures, but about encouraging more adaptability and resilient defenses.</p><h4>References</h4><p>[1] W. R. Ashby, An Introduction to Cybernetics, Chapman &amp; Hall, 1956. <a href="https://ashby.info/Ashby-Introduction-to-Cybernetics.pdf">https://ashby.info/Ashby-Introduction-to-Cybernetics.pdf</a></p><p>[2] Open Worldwide Application Security Project (OWASP), “LLM Prompt Injection Prevention Cheat Sheet,” <a href="https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html.">https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html.</a></p><p>[3] National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile,” National Institute of Standards and Technology (NIST), Gaithersburg, MD, 2024. <a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf">https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf</a></p><p>[4] Y. Liu, Y. Jia, R. Geng, J. Jia and N. Z. Gong, “Formalizing and Benchmarking Prompt Injection Attacks and Defenses,” in USENIX Association, 2024. <a href="https://www.usenix.org/conference/usenixsecurity24/presentation/liu-yupei">https://www.usenix.org/conference/usenixsecurity24/presentation/liu-yupei</a></p><p>[5] National Institute of Standards and Technology, “Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations,” National Institute of Standards and Technology, 2025. <a href="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2025.pdf">https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2025.pdf</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d891b0b7e823" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The End of Passwords? How Passkeys Are Redefining Online Security]]></title>
            <link>https://medium.com/@wakif360/the-end-of-passwords-how-passkeys-are-redefining-online-security-9a88f4e0c8c0?source=rss-6c9dd61c736f------2</link>
            <guid isPermaLink="false">https://medium.com/p/9a88f4e0c8c0</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[passkey]]></category>
            <dc:creator><![CDATA[MD Wakif Ibne Mizan]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 17:38:02 GMT</pubDate>
            <atom:updated>2026-10-15T17:38:02.961Z</atom:updated>
            <content:encoded><![CDATA[<p>Daily, millions of people log in to their various social media or online accounts. These online accounts can include online banking, e-mail, and digital services. Etc. For decades, passwords have been the only authentication method for these logins. But they are also a weak point in online security and privacy. Many issues often arise revolving around the matter of storing and remembering passwords, along with a persistent risk of data being stolen. But after decades of password-dependent authentication systems, digital advancement is leading us to an end to passwords by introducing many new authentication methods; the most discussed and trusted of them is passkeys.</p><h3>What are Passwords and Passkeys?</h3><p>A password is an arrangement of numbers, letters, and symbols with a definite length. It is like a key to a door lock. The password is unique and secret; the user has to ensure its secrecy by not sharing it. The specific arrangement of characters is stored in the system database of the software where the user has an account; he proves his identity to the system by entering the set of characters he selected as a password.<br>A passkey is a digital credential that authenticates the identity of the user when logging in without requiring him to enter a password. It can be used by a device or passkey manager, while the passkey might be a fingerprint, facial recognition, or other authentication methods.</p><h3>Understanding Cryptography</h3><p>For further understanding of passkeys, we have to go through a term called cryptography. Cryptography is a branch of cybersecurity that uses mathematical processes to ensure secure access and safer communication. It enables confidentiality of sensitive data, verifies authenticity, and restricts unauthorized access and manipulation. Cryptography holds a major role in modern digital systems, including authentication technologies. The process of verifying the validity of a digital response through a cryptographic method is called cryptographic verification.</p><h3>Understanding Passkeys</h3><p>Passkeys work with public-key cryptography; the process uses mathematical relations to create two keys: a public key and a private key. Here, the keys are credentials. The moment a user creates a passkey for a site, these two keys are generated. The public key remains in site storage, and the private key is stored on the user’s device or passkey-providing platform. The private key stays hidden.<br>When the user tries to log in to the site, it sends a unique cryptographic task to the user’s device. Then the user accepts the request, typically by device PIN, facial recognition, fingerprint, etc. Afterwards, the device generates a digital signature for the task, which is compared to the site’s stored public key; a successful match means successful authentication.<br>An important feature of passkeys is that they work for legitimate websites only, which provides security from phishing attacks. Passkeys can be bound to devices or can be used across trusted devices when provided by a passkey provider platform, while the main mechanism remains the same.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/820/1*_w29avhdQkX0xs__H4oiRA.png" /></figure><p>Passkeys provide various extra advantages over passwords in terms of security and accessibility. The phishing resistance due to their functionality being limited to legitimate websites only is one of the most important features they carry. Another massive problem that they solve is password reuse. People often use the same password on multiple sites, while with passkeys, it automatically generates unique keys for all different domains. So, if a security breach occurs, it can’t reach other accounts. Additionally, a website having passkeys only reveals the public key rather than a private credential like a password, so if a data breach occurs, the breached public data can’t give access to the user’s account. These various security and privacy features ensure enhanced security of a user’s accounts across multiple platforms.</p><p>For Individuals, the passkey can make regular tasks safer and more convenient in the digital space. For an individual user, the privacy of their stored data, like photos, videos, documents, etc., is crucial. For that purpose, passkeys can add an extra layer of security. Moreover, it will lessen the burden of remembering too many passwords since the passkey system ensures convenience. But alongside all these, a user must have an alternative account recovery option.</p><p>For commercial uses, it is undoubtedly a great integration for preserving the security of customers’ and employees’ data and accounts. In commercial use, the data and accounts managed are massive in quantity. Therefore, an extra security layer becomes a requirement. By adopting the passkey method, a business can become free of its password storage burden and the hassle of password resetting, along with the issues of stolen or reused credentials. Passkeys can also help protect admin accounts, which require maximum security. Unauthorized access to these accounts can lead to disastrous consequences. However, the successful setup may require proper infrastructure, employee training, and recovery procedures, along with backups.</p><h3>Possible downsides</h3><p>Alongside so many beneficial qualities, it surely has its limitations. The primary concern arises from the issue of recovery. If the passkey-synchronized device or account gets compromised or access is lost, it will result in a very tough recovery process. Another concern is compatibility; although support is getting broader, many older devices may not support passkeys. And finally, the passkey method may not eliminate all forms of cybersecurity threats. Some threats, like malware, social engineering, compromised devices, etc., are still effective. Therefore, rather than adopting it as a complete solution, it might be considered a preventive measure against threats.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9p281a_d3nnHGjpzhxTbHg.png" /></figure><p>The transition towards passkeys is an ongoing process right now. Many big names, like Microsoft, Amazon, PayPal, etc., have already adopted passkey methods into their systems. The shift is also reflected in recent industry data. In its October 2026 Passkey Index, the FIDO Alliance reported that passkeys accounted for 36% of successful sign-ins across the participating services, up from 26% in 2025. This suggests that passkeys are moving beyond simple availability toward regular use. Source: FIDO Alliance. But that doesn’t initiate an overnight shift; rather, it is a rising awareness. Passkeys are a step toward a secure digital future. Although it isn’t an ultimate solution, it is a step toward digital advancement.</p><h3>Further Reading</h3><p>For readers who want to explore passkeys beyond this article, the following resources offer further insights into their technology, security, implementation, and current limitations.</p><ol><li><a href="https://fidoalliance.org/passkeys/"><strong>Passkeys: Passwordless Authentication — FIDO Alliance</strong></a><br> An introduction to passkeys, their benefits, and the standards behind passwordless authentication.</li><li><a href="https://www.w3.org/TR/2026/REC-webauthn-3-20260825/"><strong>Web Authentication (WebAuthn) Level 3 — W3C</strong></a><br> The official technical specification explaining the web authentication framework that enables passkey-based authentication.</li><li><a href="https://developers.google.com/identity/passkeys"><strong>Passkeys — Google for Developers</strong></a><br> A practical technical guide covering passkey implementation, public-key cryptography, and phishing resistance.</li><li><a href="https://www.nist.gov/news-events/news/2024/04/giving-nist-sp-800-63b-boost-nist-sp-800-63b-supplement-incorporating"><strong>Giving NIST SP 800–63B a Boost: Guidance for Syncable Authenticators — NIST</strong></a><br> Explores security considerations and implementation guidance for synchronizable authenticators, including passkeys, in organizational environments.</li><li><a href="https://doi.org/10.1016/j.cose.2025.104823"><strong>Usable Authentication: Are We There Yet? — Computers &amp; Security (2026)</strong></a><br> An academic research article examining the usability of modern authentication methods, including passkeys, alongside challenges involving accessibility and cross-device experiences.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9a88f4e0c8c0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Memory Poisoning: Understanding Threats to AI Memory]]></title>
            <link>https://medium.com/@wakif360/ai-memory-poisoning-understanding-threats-to-ai-memory-2108d24d582a?source=rss-6c9dd61c736f------2</link>
            <guid isPermaLink="false">https://medium.com/p/2108d24d582a</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[MD Wakif Ibne Mizan]]></dc:creator>
            <pubDate>Wed, 14 Oct 2026 20:49:02 GMT</pubDate>
            <atom:updated>2026-10-14T20:49:02.725Z</atom:updated>
            <content:encoded><![CDATA[<p>Let’s think of a scenario in which an AI security threat is caused by a stored memory in a database. That’s what we call memory poisoning. AI is gradually becoming more than a tool. It can remember past interactions, use them to provide a personalized result in the future. But what if the stored data is manipulated? A malicious or false instruction in it can influence the future decisions it makes. So, is it a threat? Indeed it is, not just a threat but a question about trust, security, and information protection. But for now, we move on to understanding how things work.</p><p>In order to understand AI memory, we need to have a clear idea about what AI memory actually refers to. AI memory means the ability of AI to store the data of past interactions and use it to improve and personalize the results in future uses. Due to this, it doesn’t need to start the user understanding part from scratch. This becomes beneficial for someone requiring consistent use of AI. It remembers the preferences, previous instructions, the user’s general requirements, and other relevant information. So it can meet the user’s needs accordingly. Meanwhile, AI memory isn’t human memory; it can be summaries or databases that benefit the understanding of context. It makes it more useful by influencing future decisions.</p><p>Now let’s have a brief understanding of AI memory poisoning. Memory poisoning in AI refers to such activity by attackers in which they try to manipulate the information in the database to influence its future responses. The Aim is not just to manipulate a single conversation but to affect the system database to centrally impact the future responses of that AI model. The pushed information can be misleading or even malicious. Imagine it as an AI model working on a project; if an attacker succeeds in memory poisoning it, the AI model might take some destructive measures regarding the project, where the intensity of impact depends on the level of access and permissions provided. AI systems rely heavily on stored memories, which makes it more necessary to protect their stored data. This issue concerns all users.</p><p>AI memory poisoning can be of many types, depending on the attack and impact. The four most important types are mentioned.</p><p><strong>False Information Injection:</strong> When an attacker inputs misleading information into the AI model to affect its future responses.</p><p><strong>Malicious Instruction Injection:</strong> Harmful malicious injection code can be input into the AI database, which can cause harm to users in the future.</p><p><strong>Memory manipulation through user interaction:</strong> An attacker may influence an AI about what it remembers through carefully constructed conversations.</p><p><strong>Cross-Task and Cross-User poisoning:</strong> When malicious information injected in one task or user context affects another task or user context due to improper isolation of data.</p><p>Other possible approaches include poisoning stored summaries, manipulating retrieved documents, and exploiting weaknesses in memory-update or retrieval processes. These categories can overlap, and their feasibility depends on the AI system’s memory architecture and security controls.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WpU_wim7w0HBUSti25YQLw.png" /></figure><p>Focusing on the effects of memory poisoning, we first need to talk about the risks in single-user or personal use. The risk in single-user settings often goes unnoticed. This is often related to personal advice or daily planning by an AI assistant. It affects the user with inaccurate recommendations, false or misleading advice, and even unreal data. The risk is more concerning when the user completely relies on the AI model. They put too much trust in it, which later proves harmful. It is a concern for personal use.</p><p>In commercial uses of AI, the effects can be beyond just inaccurate answers and include more concerning risks. For commercial use and businesses today, AI handles customer support, planning, automated workflow, research, and software development. If somehow the memory is compromised by poisoning through malicious injection, it may even leak sensitive data. It can even interrupt the routine workflow. For instance, an AI agent can cause incorrect or buggy code in software; a customer service chatbot can mislead users with false information, etc. The consequences also depend on how much data exposure and permissions are provided to the AI model. Proper safeguards in the commercial use of AI are a matter of concern. It is related to the company KPI, financial loss or profit, growth, customer engagement, etc.</p><p>Now, to have a clearer concept, let’s understand the impact of memory poisoning in AI chatbots and AI agents. In the case of an AI chatbot, due to memory poisoning, it may affect the answers, the information, or the recommendations. The answers from the chatbot may confuse or mislead the user. In the case of AI agents, the effect is more severe. Since agentic AI is responsible for advanced automation, a simple hazard can be extremely harmful. It contains connections with multiple software, files, and databases, elevating the risk to another level. Unauthorized actions, improper results, and inappropriate outcomes can lead to significant harm in commercial tasks.</p><p>In order to prevent memory poisoning in AI, data filtering must be made strong by safeguarding the data storage process. One of the important measures for prevention is “Memory Validation”, where the system is designed to assess the reliability, security, and sources of information before using it for future responses. Another important measure is to audit memory at regular intervals, so that harmful or outdated information can be removed.</p><p>For users, they should be careful about what type of solution an AI model is providing to them; it must be verified before taking any actions, and in automation, one should stay concerned about what level of data access one is allowing to the AI agent. Users must be aware of AI safeguards.</p><p>Undoubtedly, AI systems make task completion more efficient. But along with advancing, various security risks keep arising, and safeguarding becomes more critical. Memory poisoning highlights how input data can affect the future decisions of an AI model. But the goal must not be to make AI forget everything or stop using automation. The goal must be ensuring smart use of AI automations. After all, building trustworthy AI is not just about what these systems can learn or accomplish; it is also about ensuring that what they remember does not become a threat.</p><h3>Further Reading</h3><p>For readers interested in exploring AI memory poisoning in greater depth, the following research papers provide additional insights into memory vulnerabilities, persistent attacks, and potential defense mechanisms.</p><ol><li><strong>From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents</strong><br>Link: <a href="https://arxiv.org/abs/2606.04329">https://arxiv.org/abs/2606.04329</a></li><li><strong>MemSecBench: Tracking Agent Memory Poisoning from Persistence to Consequence and Repair</strong><br>Link: <a href="https://arxiv.org/abs/2607.27080">https://arxiv.org/abs/2607.27080</a></li><li><strong>Memory Poisoning Attack and Defense on Memory Based LLM-Agents</strong><br>Link: <a href="https://arxiv.org/abs/2601.05504">https://arxiv.org/abs/2601.05504</a></li><li><strong>PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models</strong><br>Link: <a href="https://www.usenix.org/conference/usenixsecurity25/presentation/zou-poisonedrag">https://www.usenix.org/conference/usenixsecurity25/presentation/zou-poisonedrag</a></li><li><strong>MemSentry: A Framework for Detecting Persistent Memory Poisoning in Agentic AI</strong><br>Link: <a href="https://arxiv.org/abs/2609.08747">https://arxiv.org/abs/2609.08747</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2108d24d582a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Chatbots to Autonomous Agents: The Big Shift in AI]]></title>
            <link>https://medium.com/@wakif360/from-chatbots-to-autonomous-agents-the-big-shift-in-ai-1bb7ddc1dfb4?source=rss-6c9dd61c736f------2</link>
            <guid isPermaLink="false">https://medium.com/p/1bb7ddc1dfb4</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[MD Wakif Ibne Mizan]]></dc:creator>
            <pubDate>Tue, 13 Oct 2026 09:03:58 GMT</pubDate>
            <atom:updated>2026-10-13T09:03:58.347Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i5rZ09ZG2v_nYBLtctznJw.png" /></figure><p>For a long time, AI has been trained to respond with a solution according to our needs. Now we are witnessing a gradual shift toward AI taking action. Yes, I mean autonomous AI agents. The newest form of AI that thinks, decides, takes action, and brings results. Undoubtedly, it can be considered a drastic change in the AI usage process and the mentality of people towards AI.</p><p>Let’s have a look at what an AI Agent is. First, we think about our conventional AI chatbots. They are the kind of AI that works like a smart consultant. You come up with a problem to get required or approximate solutions. It answers questions, gives a solution, and lets you take further actions. But AI agents work differently. They work as your personal assistant, who not only gives a solution but also takes action and follows your commands, completing a workflow. The major difference here can be considered the capability of taking actions.</p><p>The major shift is emerging rapidly. This shift was initiated by integrating an AI model into an application so that it can understand context, make the best decision, and take actions. This is widely used now, but along with this, a new form of agentic AI has already emerged. Advanced AI agents can take complete actions according to your needs. It brings up a completely new era of personal and commercial autonomy. Autonomy in AI agents exists on a spectrum. A fully autonomous agent combines four capabilities in a single system:</p><p><strong>Goal decomposition:</strong> breaking a broad objective into concrete subtasks.</p><p><strong>Tool invocation:</strong> autonomously deciding when and how to use external systems, browsers, code interpreters, APIs, file systems, and databases.</p><p><strong>Self-evaluation and recovery:</strong> checking its own output against the goal, detecting failures, and retrying with a different approach.</p><p><strong>Long-horizon persistence:</strong> maintaining context and memory across hours or days of work, not just a single conversation turn.</p><p>Traditional automation follows a script written by a human. Autonomous agents write their own plan at runtime and adapt it as conditions change.</p><p>This major shift is facilitated by multiple factors. The major impact is driven by new and advanced LLMs, which are rapidly improving day by day. Along with this, an AI model becomes more practically useful when it can interact with external tools. Depending on the permissions it receives, an agent may search the web, read documents, execute code, query databases, or interact with software through APIs. Agents may use conversation history, stored information, or external databases to maintain context across multiple steps in order to help with remembering and tracking context and process. Picture directing an AI assistant to investigate a subject. A traditional chatbot would likely offer a summary along with links to relevant resources. An agentic system, on the other hand, could go further: gathering information from numerous sources, arranging the material, evaluating contradictory claims against each other, assembling a well-organized report, and storing the finished product in a specified location.</p><p>Giving AI access to advanced autonomy might be an efficient approach. But the security downsides should also be pointed out. Let us bring up some points to have a clear vision of the downsides.</p><p><strong>Prompt Injection:</strong> These systems interpret natural language rather than following rigid code rules; they can be manipulated by malicious content embedded in ordinary data sources like webpages or emails. This risk is amplified when agents have access to sensitive tools or private files.</p><p><strong>Excessive access:</strong> A greater extent of access is to be provided to AI for better autonomy. But still, there are privacy concerns, so it must be kept in mind that the data given access to must not be anything confidential. Along with it, unrestricted access must be supervised.</p><p><strong>Agent Hijacking:</strong> An AI agent being used for commercial use has to be protected intensively. A compromised AI agent is not something out of the box; instead, it is a notable risk factor.</p><p>Along with these, there are many other severe limiting factors: accountability gap, cascading failure, data leakage, etc.</p><p>But in a broader view, we should consider this shift as an opportunity. With ethical development and use of these agentic autonomous AI systems, various repetitive and complex tasks will become easier to accomplish. A huge time saving will lead to a faster development of technology. Ultimately, its greatest advantage may not be replacing human intelligence, but extending what humans can accomplish when powerful AI systems operate as capable, well-supervised partners.</p><p>The development of agentic artificial intelligence indicates a fundamental change in how individuals interact with computational systems. Instead of software that only provides information, those programs now execute specific actions for the user. The capacity for these tools to change commercial sectors, organize complicated processes, and increase the skills of developers is clear — but as users give those systems more independence, the software creates vulnerabilities in digital security and makes legal responsibility difficult to define. The primary difficulty is not just the creation of more powerful agents. The goal is to ensure that their actions match the goals of the user plus function within defined limits. In reality the significant “AI takeover” is likely the current trend where users slowly transfer their choices and duties to the programs. There is a question regarding the number of tasks that users ought to delegate to artificial intelligence. And it is uncertain if society is ready for the outcomes of this transfer.</p><p><strong>Further Reading</strong></p><p>The following resources offer additional insights into agentic AI, its security challenges, and the risks associated with increasingly autonomous systems.</p><p><strong>OWASP GenAI Security Project:</strong> An overview of security risks in generative AI applications, including prompt injection, sensitive information disclosure, and excessive agency. <a href="https://genai.owasp.org/">https://genai.owasp.org/</a></p><p><strong>NIST AI Risk Management Framework:</strong> A framework for understanding, assessing, and managing risks associated with the development and use of AI systems. <a href="https://www.nist.gov/itl/ai-risk-management-framework">https://www.nist.gov/itl/ai-risk-management-framework</a></p><p><strong>MITRE ATLAS:</strong> A knowledge base documenting adversarial tactics and techniques that target AI-enabled systems, useful for exploring AI-related cybersecurity threats. <a href="https://atlas.mitre.org/">https://atlas.mitre.org/</a></p><p><strong>OWASP Top 10 for Large Language Model Applications:</strong> Guidance on major security vulnerabilities affecting LLM-powered applications and potential approaches to mitigating them. <a href="https://genai.owasp.org/llm-top-10/">https://genai.owasp.org/llm-top-10/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1bb7ddc1dfb4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>