<?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 Myojin Research on Medium]]></title>
        <description><![CDATA[Stories by Myojin Research on Medium]]></description>
        <link>https://medium.com/@Myojin.research?source=rss-93eac886451b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*tVIr9BYvLkMgEUvv0XZ9Rw.png</url>
            <title>Stories by Myojin Research on Medium</title>
            <link>https://medium.com/@Myojin.research?source=rss-93eac886451b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 06:22:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Myojin.research/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 Doesn’t Have to Lie to Change Your Past]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/fractal-cognition-layer/ai-doesnt-have-to-lie-to-change-your-past-4ed7a952d885?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1672/1*o6r4xCtgLymLPBoag5LQUQ.png" width="1672"></a></p><p class="medium-feed-snippet">When artificial intelligence begins telling the story of your life, it may also influence the person you are becoming.</p><p class="medium-feed-link"><a href="https://medium.com/fractal-cognition-layer/ai-doesnt-have-to-lie-to-change-your-past-4ed7a952d885?source=rss-93eac886451b------2">Continue reading on Fractal Cognition Layer »</a></p></div>]]></description>
            <link>https://medium.com/fractal-cognition-layer/ai-doesnt-have-to-lie-to-change-your-past-4ed7a952d885?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ed7a952d885</guid>
            <category><![CDATA[self-awareness]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[psychology]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 20:22:20 GMT</pubDate>
            <atom:updated>2026-11-01T20:22:20.814Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[What Are We Actually Calling Secure? Lessons From PHP2Shell]]></title>
            <link>https://ai.plainenglish.io/what-are-we-actually-calling-secure-lessons-from-php2shell-ff370b951db8?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff370b951db8</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ethical-hacking]]></category>
            <category><![CDATA[application-security]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[php]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Sun, 25 Oct 2026 04:12:11 GMT</pubDate>
            <atom:updated>2026-10-25T04:17:05.860Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J6nrZEYnhc7V-Z7RBTy7Vw.png" /></figure><p><em>Before declaring a system safe, ethical hackers should ask a different question: What assumptions make that judgment possible?</em></p><p>On October 21, 2026, the PHP development team disclosed two critical vulnerabilities: CVE-2026–109618 and CVE-2026–109619.</p><p>Collectively known as PHP2Shell, these flaws can allow unauthenticated remote code execution under vulnerable conditions.</p><p>The PHP team released patched versions and urged users to update immediately.</p><p>But one detail is particularly important.</p><p>The official advisory describes an attack path in which a request can trigger exploitation before any application code runs.</p><p>Think about what that means.</p><p>Your authentication logic might be correct.</p><p>Your authorization checks might work exactly as intended.</p><p>Your application might pass its security tests.</p><p>And yet, the system could still be vulnerable through the runtime beneath it.</p><p>This raises a question that deserves more attention:</p><p><strong>When we say a system is secure, what exactly have we verified?</strong></p><h3>A passed security test is not proof of a secure system</h3><p>Security assessments necessarily work within boundaries.</p><p>A penetration tester may examine authentication, session management, access controls, input validation, and business logic.</p><p>These are essential activities.</p><p>But every assessment also depends on assumptions about the environment being tested.</p><p>Consider a PHP application with a well-designed authentication system.</p><p>A security review finds no obvious authorization bypass. User roles are enforced correctly. Sensitive endpoints reject unauthorized requests.</p><p>Within that scope, the findings are encouraging.</p><p>But what has actually been established?</p><p>The application appears to enforce its intended access controls under the tested conditions.</p><p>That conclusion does not establish that the PHP runtime is free from memory-corruption vulnerabilities.</p><p>It does not establish that every deployed package has received the appropriate security patches.</p><p>And it certainly does not establish that the system was never compromised before the assessment.</p><p>These are different claims requiring different evidence.</p><p>The problem is not necessarily that the original assessment was wrong.</p><p><strong>The problem begins when a valid conclusion is extended beyond the evidence supporting it.</strong></p><p>For ethical hackers, recognizing that boundary can be as important as identifying a vulnerability.</p><h3>What PHP2Shell reveals</h3><p>PHP2Shell provides a concrete example.</p><p>According to the PHP security advisory, the two vulnerabilities involve problems with internal request-memory handling.</p><p>CVE-2026–109618 involves out-of-memory handling and partially initialized internal state.</p><p>CVE-2026–109619 involves internal state that can continue referring to memory after it has been freed.</p><p>Both may lead to remote code execution under vulnerable conditions.</p><p>The security team specifically highlighted web-facing PHP installations, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP, and LiteSpeed.</p><p>For mod_php, the advisory described a demonstrated unauthenticated attack path that can be triggered before application code executes.</p><p>This matters because it exposes a separation between two kinds of assurance.</p><p>One concerns the application.</p><p>The other concerns the environment that makes the application possible.</p><p>A secure login implementation cannot compensate for a vulnerable underlying runtime.</p><p>Nor can a functioning web application firewall be treated as proof that a vulnerable PHP installation has been fully protected.</p><p>The official PHP guidance states that updating is the only complete solution for these vulnerabilities.</p><p>As of October 25, 2026, the relevant fixed PHP releases are:</p><p>PHP branchFixed releasePHP 8.28.2.35PHP 8.38.3.36PHP 8.48.4.27PHP 8.58.5.12</p><p>Vendor-maintained packages may contain backported fixes, so an effective assessment must also consider distribution-specific advisories and package builds.</p><p>There is another distinction worth preserving.</p><p><strong>A system that has been patched is not necessarily a system that has been cleared of prior compromise.</strong></p><p>Patching addresses the vulnerability. Investigating earlier exposure, suspicious activity, and possible persistence requires separate evidence.</p><h3>The assumptions worth investigating</h3><p>Imagine being asked to review a production web service.</p><p>You receive documentation showing that:</p><p>The application has passed its authentication tests.</p><p>A WAF is active.</p><p>The development team has deployed recent application changes.</p><p>The server is reported to be running a supported PHP release.</p><p>Nothing in this description necessarily indicates an active compromise.</p><p>But several questions remain unanswered.</p><p>Which PHP binary is actually processing incoming requests?</p><p>Is the installed package patched, including any relevant vendor backports?</p><p>Are containers, managed hosting services, and background workers using the same runtime?</p><p>Have all affected web-facing environments been identified?</p><p>Was the system exposed before the fix was applied?</p><p>What evidence is available to investigate that period?</p><p>These questions do not establish that a vulnerability exists.</p><p>They identify what remains necessary before a broader security judgment can be justified.</p><p>That difference is central to responsible security work.</p><p>A missing piece of evidence is not automatically a finding.</p><p>But it should not quietly become evidence of safety, either.</p><h3>A practical AI prompt for ethical hackers</h3><p>AI can help review large amounts of security documentation, source code, deployment information, and vulnerability advisories.</p><p>It can also make mistakes.</p><p>An AI assistant may generate a technically plausible security report while assuming that the supplied system boundary is complete.</p><p>It may treat a software version string as sufficient evidence of patch status.</p><p>Or it may propose an unnecessary test because it has not recognized that the available evidence already resolves the question.</p><p>The following prompt is designed to make those assumptions more visible.</p><p>It is intended for authorized security reviews and defensive analysis.</p><p>It does not replace penetration testing, threat modeling, incident response, or independent technical verification.</p><h3>Security Assumption Review Prompt</h3><p>Copy the complete prompt below into your AI assistant, then replace the placeholders with information from your authorized assessment environment. No exploitation or unauthorized access is required.</p><pre>ROLE</pre><pre>Act as a security review assistant supporting<br>an authorized ethical hacker or defensive<br>security team.</pre><pre>Your task is not merely to enumerate<br>vulnerabilities.</pre><pre>Examine what is actually known about the<br>system, which security conclusions the<br>available evidence supports, and which<br>conclusions depend on unverified assumptions.</pre><pre>SCOPE</pre><pre>System or application:<br>[Describe the system]</pre><pre>Authorized assessment scope:<br>[Define permitted systems and activities]</pre><pre>Available evidence:<br>[Architecture, configuration, advisories,<br>version information, existing test results]</pre><pre>Known constraints:<br>[Production limitations, unavailable logs,<br>third-party dependencies, prohibited actions]</pre><pre>REVIEW OBJECTIVES</pre><pre>1. Identify relevant observable behavior,<br>   components, dependencies, permissions,<br>   and trust boundaries.</pre><pre>2. State the evidence supporting each<br>   existing security conclusion.</pre><pre>3. Identify assumptions required for those<br>   conclusions to remain valid.</pre><pre>4. Examine whether those assumptions change<br>   across software layers, services,<br>   deployment environments, or time.</pre><pre>5. Distinguish application-level controls<br>   from the security of underlying runtimes,<br>   libraries, infrastructure, and vendors.</pre><pre>6. Consider whether individually correct<br>   components could produce unsafe outcomes<br>   through their interactions.</pre><pre>7. Identify missing evidence without<br>   automatically treating uncertainty<br>   as proof of vulnerability.</pre><pre>8. If the available evidence shows that<br>   the current assessment scope is<br>   insufficient, explain the missing<br>   dependency and recommend an appropriately<br>   authorized expansion of the review.</pre><pre>9. For each potential issue, distinguish:<br>   - Confirmed facts<br>   - Supported inferences<br>   - Conditional risks<br>   - Unverified hypotheses<br>   - Unknowns</pre><pre>10. Recommend proportionate, non-destructive<br>    verification methods and defensive<br>    improvements.</pre><pre>EVIDENCE RULES</pre><pre>Do not invent vulnerabilities.</pre><pre>Do not assume that a passed test establishes<br>the security of the entire system.</pre><pre>Do not treat a patch as proof that no<br>earlier compromise occurred.</pre><pre>Verify time-sensitive vulnerability claims<br>against authoritative vendor advisories.</pre><pre>Do not confuse missing evidence with<br>evidence of exploitation.</pre><pre>Do not recommend unauthorized access,<br>destructive testing, or production<br>exploitation.</pre><pre>Do not submit secrets, private keys,<br>credentials, or sensitive customer data<br>to external services.</pre><pre>OUTPUT FORMAT</pre><pre>A. Verified security facts</pre><pre>B. Existing security assumptions</pre><pre>C. Unverified dependencies</pre><pre>D. Potential failure conditions</pre><pre>E. Recommended safe verification</pre><pre>F. Remediation or risk-reduction options</pre><pre>G. Remaining unknowns</pre><pre>H. Decisions that cannot yet be justified</pre><pre>Prioritize findings that could materially<br>change a real security decision.</pre><pre>Avoid speculative vulnerability lists<br>that have no connection to the evidence.</pre><h3>What this prompt is meant to improve</h3><p>The purpose is not to make an AI sound more sophisticated.</p><p>It is to produce a more useful assessment.</p><p>A useful review should help the analyst understand what is known, what remains uncertain, and which missing fact could change the security decision.</p><p>For example, when reviewing an environment affected by PHP2Shell, the assistant might identify that application-layer testing cannot determine runtime patch status.</p><p>It might recommend checking deployment artifacts or the appropriate vendor advisory.</p><p>It might identify a need to distinguish remediation status from historical compromise assessment.</p><p>Those are actionable observations.</p><p>Whether the prompt improves real-world security assessments compared with an ordinary AI workflow remains an empirical question. It should be tested, not assumed.</p><h3>The next security boundary may be the assessment itself</h3><p>There is another problem during fast-moving vulnerability disclosures.</p><p>A security team discovers an urgent advisory.</p><p>The team searches for a diagnostic tool.</p><p>An AI assistant suggests remediation commands.</p><p>Engineers begin making changes under time pressure.</p><p>Each action is intended to reduce risk.</p><p>But the response process has its own security dependencies.</p><p>Is the advisory authentic?</p><p>Is the diagnostic utility trustworthy?</p><p>Does the proposed remediation match the actual deployment?</p><p>Could a configuration change interrupt essential services?</p><p>Are logs and forensic evidence being preserved where appropriate?</p><p>The effort to restore security can introduce new risks if its own assumptions are left unexamined.</p><p>This does not mean teams should delay critical patching until every question has been answered.</p><p>It means verified emergency actions and unresolved questions must remain distinct.</p><p>The goal is to move quickly without losing track of what has — and has not — been established.</p><h3>From vulnerability research to organizational resilience</h3><p>For an ethical hacker, this perspective can help locate overlooked assessment boundaries.</p><p>For a security engineer, it can reveal weaknesses in deployment assumptions and remediation processes.</p><p>For an organization, it becomes a question of dependency.</p><p>A company may control its application code but depend on cloud providers, operating systems, open-source components, managed services, and external vendors.</p><p>Its effective security posture depends partly on how those relationships are maintained and verified.</p><p>The same problem extends to critical infrastructure.</p><p>Healthcare, financial services, telecommunications, and public administration often depend on software components shared across many organizations.</p><p>A serious vulnerability in a common dependency can therefore create correlated risks, even when individual organizations have implemented their own security controls correctly.</p><p>For individuals, the implications are quieter but still significant.</p><p>The safety of a digital account depends not only on password practices or multifactor authentication, but also on systems the account holder cannot directly inspect.</p><p>These are not identical security problems.</p><p>Yet each involves a difference between the security measures that are visible and the conditions on which actual safety depends.</p><h3>What should AI-assisted security become better at?</h3><p>AI is becoming increasingly capable of writing code, inspecting configurations, summarizing advisories, and generating security tests.</p><p>Those capabilities can be valuable.</p><p>But generating more findings is not always the same as improving an assessment.</p><p>Sometimes a critical contribution is recognizing that the current evidence does not support the conclusion being drawn.</p><p>Sometimes it is discovering that the scope of investigation was too narrow.</p><p>And sometimes it is recognizing that the question being investigated is no longer the one that matters.</p><p>A capable security assistant should not merely produce answers within a given frame.</p><p>It should also help an authorized analyst understand when that frame no longer fits the evidence.</p><p>That is a useful direction for AI-assisted cybersecurity — one that should be judged through reproducible results rather than impressive language.</p><h3>The question to leave open</h3><p>A vulnerability is something a security researcher may discover.</p><p>But there is another kind of discovery.</p><p>A system was believed to be safe.</p><p>Its components appeared to work correctly.</p><p>Its tests returned reassuring results.</p><p>Then a previously unexamined dependency changed what those results actually meant.</p><p>The most valuable finding may not be another entry on a vulnerability checklist.</p><p>It may be the assumption that made the system appear secure in the first place.</p><p><strong>Before asking whether a system is secure, ask what would have to be true for that conclusion to hold.</strong></p><p>And then find out which of those conditions have actually been verified.</p><h3>References</h3><ul><li><a href="https://www.php.net/archive/2026.php">PHP Security Advisory — October 21, 2026</a></li><li><a href="https://www.ipa.go.jp/security/security-alert/2026/alert20261022.html">IPA: PHP2Shell Vulnerability Advisory</a></li><li><a href="https://cheatsheetseries.owasp.org/cheatsheets/Threat_Modeling_Cheat_Sheet.html">OWASP Threat Modeling Cheat Sheet</a></li><li><a href="https://owasp.org/www-project-web-security-testing-guide/">OWASP Web Security Testing Guide</a></li></ul><p><em>This article presents an exploratory application of security-assumption analysis informed by MyojinModel research. It does not disclose the model’s internal methods or claim a demonstrated advantage over established security practices. The included prompt is intended for authorized, defensive use.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff370b951db8" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/what-are-we-actually-calling-secure-lessons-from-php2shell-ff370b951db8">What Are We Actually Calling Secure? Lessons From PHP2Shell</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Can Act Before We Understand. But Who Decides?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://ai.plainenglish.io/what-happens-when-expertise-is-no-longer-required-before-action-7832f763ba81?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1920/1*t62OD4nXVKfAXevnlwJG3Q.png" width="1920"></a></p><p class="medium-feed-snippet">The October 2026 security crisis reveals a deeper shift in human expertise &#x2014; and why the ability to act is not the same as the ability to&#x2026;</p><p class="medium-feed-link"><a href="https://ai.plainenglish.io/what-happens-when-expertise-is-no-longer-required-before-action-7832f763ba81?source=rss-93eac886451b------2">Continue reading on Artificial Intelligence in Plain English »</a></p></div>]]></description>
            <link>https://ai.plainenglish.io/what-happens-when-expertise-is-no-longer-required-before-action-7832f763ba81?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/7832f763ba81</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[human-intelligence]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[critical-thinking]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 14:17:41 GMT</pubDate>
            <atom:updated>2026-10-23T15:04:10.879Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI May Change the Reason a Task Exists]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/fractal-cognition-layer/ai-may-change-the-reason-a-task-exists-e7f4da848ff6?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1672/1*QfQQ_k_pMtAVr1I_d7jY7A.png" width="1672"></a></p><p class="medium-feed-snippet">As AI agents become capable of doing more work, some of the assumptions behind that work may begin to disappear.</p><p class="medium-feed-link"><a href="https://medium.com/fractal-cognition-layer/ai-may-change-the-reason-a-task-exists-e7f4da848ff6?source=rss-93eac886451b------2">Continue reading on Fractal Cognition Layer »</a></p></div>]]></description>
            <link>https://medium.com/fractal-cognition-layer/ai-may-change-the-reason-a-task-exists-e7f4da848ff6?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e7f4da848ff6</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 15:52:18 GMT</pubDate>
            <atom:updated>2026-10-19T15:52:18.558Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Where Understanding Used to Live:]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://ai.plainenglish.io/where-understanding-used-to-live-84b2d979077c?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1672/1*65bZG-pw50X5cZcirUVPQQ.png" width="1672"></a></p><p class="medium-feed-snippet">How AI Is Quietly Removing the Need for Persistent Internal State</p><p class="medium-feed-link"><a href="https://ai.plainenglish.io/where-understanding-used-to-live-84b2d979077c?source=rss-93eac886451b------2">Continue reading on Artificial Intelligence in Plain English »</a></p></div>]]></description>
            <link>https://ai.plainenglish.io/where-understanding-used-to-live-84b2d979077c?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/84b2d979077c</guid>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[humancomputer-interaction]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Sat, 10 Oct 2026 21:50:37 GMT</pubDate>
            <atom:updated>2026-10-10T21:50:37.193Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The AI Safety Failure That Doesn’t Look Like a Refusal]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://ai.plainenglish.io/the-ai-safety-failure-that-doesnt-look-like-a-refusal-c0fa6ea7db86?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/2172/1*3EftreXMztnzLBBKRIBTsQ.png" width="2172"></a></p><p class="medium-feed-snippet">Some apparent limits of AI may come from how a system interprets its task, rather than what it is capable of doing.</p><p class="medium-feed-link"><a href="https://ai.plainenglish.io/the-ai-safety-failure-that-doesnt-look-like-a-refusal-c0fa6ea7db86?source=rss-93eac886451b------2">Continue reading on Artificial Intelligence in Plain English »</a></p></div>]]></description>
            <link>https://ai.plainenglish.io/the-ai-safety-failure-that-doesnt-look-like-a-refusal-c0fa6ea7db86?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c0fa6ea7db86</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[ai-research]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 14:48:20 GMT</pubDate>
            <atom:updated>2026-10-08T14:48:20.398Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Once a Frontier Model Is Public, It Starts Teaching the Competition]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@Myojin.research/once-a-frontier-model-is-public-it-starts-teaching-the-competition-370fd1be8264?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1536/1*gULNcZr7qF4nMwVK4g4w4g.png" width="1536"></a></p><p class="medium-feed-snippet">OpenAI&#x2019;s latest distillation incident raises a question that goes beyond security: when should a frontier capability become public?</p><p class="medium-feed-link"><a href="https://medium.com/@Myojin.research/once-a-frontier-model-is-public-it-starts-teaching-the-competition-370fd1be8264?source=rss-93eac886451b------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@Myojin.research/once-a-frontier-model-is-public-it-starts-teaching-the-competition-370fd1be8264?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/370fd1be8264</guid>
            <category><![CDATA[ai-strategy]]></category>
            <category><![CDATA[model-distillation]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-security]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 12:45:26 GMT</pubDate>
            <atom:updated>2026-10-08T12:45:26.846Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Once a Frontier Model Is Queryable, It Starts Teaching the Competition]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@Myojin.research/once-a-frontier-model-is-queryable-it-starts-teaching-the-competition-c3e0a59c8c99?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1672/1*Fs8vhqFHXq0Ie8pFLkaPzg.png" width="1672"></a></p><p class="medium-feed-snippet">OpenAI&#x2019;s latest distillation incident suggests that release changes more than access: it can reduce how much competitors need to discover&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@Myojin.research/once-a-frontier-model-is-queryable-it-starts-teaching-the-competition-c3e0a59c8c99?source=rss-93eac886451b------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@Myojin.research/once-a-frontier-model-is-queryable-it-starts-teaching-the-competition-c3e0a59c8c99?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c3e0a59c8c99</guid>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai-security]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 23:17:39 GMT</pubDate>
            <atom:updated>2026-10-02T23:17:39.551Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[When AI Disagrees With You, Who Gets the Final Word?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://ai.plainenglish.io/when-ai-disagrees-with-you-who-gets-the-final-word-30ae721b7da4?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1672/1*Ek90O67Bb0WXz0Ht6o9X1g.png" width="1672"></a></p><p class="medium-feed-link"><a href="https://ai.plainenglish.io/when-ai-disagrees-with-you-who-gets-the-final-word-30ae721b7da4?source=rss-93eac886451b------2">Continue reading on Artificial Intelligence in Plain English »</a></p></div>]]></description>
            <link>https://ai.plainenglish.io/when-ai-disagrees-with-you-who-gets-the-final-word-30ae721b7da4?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/30ae721b7da4</guid>
            <category><![CDATA[future-of-ai]]></category>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[human-ai-collaboration]]></category>
            <category><![CDATA[critical-thinking]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 17:49:51 GMT</pubDate>
            <atom:updated>2026-10-01T15:16:58.327Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[NVIDIA’s New Agent Safety Architecture Moves Safety Outside the AI]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@Myojin.research/ai-safety-is-moving-outside-the-ai-1031d8be86f2?source=rss-93eac886451b------2"><img src="https://cdn-images-1.medium.com/max/1536/1*t8msololpfHEvf5soMd2zg.png" width="1536"></a></p><p class="medium-feed-snippet">NVIDIA has introduced a new safety platform for autonomous AI agents. The immediate story is straightforward: as agents gain access to&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@Myojin.research/ai-safety-is-moving-outside-the-ai-1031d8be86f2?source=rss-93eac886451b------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@Myojin.research/ai-safety-is-moving-outside-the-ai-1031d8be86f2?source=rss-93eac886451b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1031d8be86f2</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[ai-safety]]></category>
            <dc:creator><![CDATA[Myojin Research]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 12:38:33 GMT</pubDate>
            <atom:updated>2026-09-29T06:50:10.365Z</atom:updated>
        </item>
    </channel>
</rss>