<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:media="http://search.yahoo.com/mrss/"><channel><title>Github Actions | FirewallSync</title><link>https://firewallsync.com/tags/github-actions/</link><description>Practical cybersecurity writing for engineering teams that ship fast.</description><language>en-us</language><image><url>https://firewallsync.com/icon-512.png</url><title>FirewallSync</title><link>https://firewallsync.com/</link></image><lastBuildDate>Thu, 19 Nov 2026 04:25:00 +0000</lastBuildDate><atom:link href="https://firewallsync.com/tags/github-actions/index.xml" rel="self" type="application/rss+xml"/><item><title>Microsoft agent-framework PyPI Compromise: What We Know</title><link>https://firewallsync.com/posts/microsoft-agent-framework-pypi-compromise-1-26-1-1-26-2-what-we-know/</link><guid>https://firewallsync.com/posts/microsoft-agent-framework-pypi-compromise-1-26-1-1-26-2-what-we-know/</guid><pubDate>Thu, 19 Nov 2026 04:25:00 +0000</pubDate><dc:creator>John Marshall</dc:creator><category>AI Security</category><description>Two malicious releases of Microsoft's agent-framework Python package, 1.26.1 and 1.26.2, were on PyPI for under an hour on November 17, 2026. Both installed a .pth file that ran a credential stealer on every Python start. What is confirmed, what is not, and how to check your environments.</description><enclosure url="https://firewallsync.com/posts/microsoft-agent-framework-pypi-compromise-1-26-1-1-26-2-what-we-know/cover_hu7113213060184927081.jpg" type="image/jpeg" length="0"/><media:content url="https://firewallsync.com/posts/microsoft-agent-framework-pypi-compromise-1-26-1-1-26-2-what-we-know/cover_hu7113213060184927081.jpg" medium="image" type="image/jpeg" width="1200" height="630"/><content:encoded><![CDATA[<p><img src="https://firewallsync.com/posts/microsoft-agent-framework-pypi-compromise-1-26-1-1-26-2-what-we-know/cover_hu7113213060184927081.jpg" alt="A heap of brown cardboard boxes and open cartons stacked against wooden shelving in a dim storeroom" width="1200" height="630"></p><p><strong>What happened:</strong> On November 17, 2026, two releases of <code>agent-framework</code>, the Python package for Microsoft Agent Framework, were published to PyPI with a credential stealer inside. Versions 1.26.1 and 1.26.2 were available for under an hour, according to <a href="https://www.wiz.io/blog/malicious-version-of-microsofts-agent-framework-published-to-pypi">Wiz</a>. Both installed a <code>.pth</code> file that ran the stealer whenever Python started in that environment, and it collected cloud, CI and registry credentials. <a href="https://www.securityweek.com/microsoft-agent-framework-releases-poisoned-with-credential-stealer/">SecurityWeek</a> reports that an Agent Framework maintainer confirmed on GitHub that the releases were removed and that an investigation is underway.</p>
<p>The package is not niche: SecurityWeek puts it at about 190,000 downloads a month. Any environment that resolved <code>agent-framework</code> without a pin in that window, on a developer laptop, a CI runner or a container build, may have installed one of the bad versions.</p>
<h2 id="timeline-utc">Timeline (UTC)</h2>
<table>
  <thead>
      <tr>
          <th>Time</th>
          <th>Event</th>
          <th>Source</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td>Nov 12, 18:25</td>
          <td>Legitimate 1.26.0 published</td>
          <td>Wiz; PyPI upload time 18:24:57 to 18:25:37</td>
      </tr>
      <tr>
          <td>Nov 17, 20:35:20</td>
          <td>Malicious 1.26.1 published</td>
          <td>Chainguard (Wiz: 20:35)</td>
      </tr>
      <tr>
          <td>Nov 17, 21:08:39</td>
          <td>Malicious 1.26.2 published, after PyPI had quarantined 1.26.1</td>
          <td>Chainguard (Wiz: about 21:08)</td>
      </tr>
      <tr>
          <td>Nov 17, about 21:30</td>
          <td>Both versions no longer available on PyPI</td>
          <td>Wiz</td>
      </tr>
      <tr>
          <td>Nov 19, about 05:00</td>
          <td>PyPI&rsquo;s JSON API lists 1.26.0 as the latest version; 1.26.1 and 1.26.2 are absent</td>
          <td>Our check of <code>pypi.org/pypi/agent-framework/json</code></td>
      </tr>
  </tbody>
</table>
<p><a href="https://www.chainguard.dev/unchained/microsoft-agent-framework-compromised-on-pypi-november-2026">Chainguard</a> says its Chainguard Libraries mirror blocked 1.26.1 48 seconds after publication and 1.26.2 42 seconds after.</p>
<h2 id="confirmed-vs-not-confirmed">Confirmed vs not confirmed</h2>
<table>
  <thead>
      <tr>
          <th>Claim</th>
          <th>Status</th>
          <th>Who says so</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td>1.26.1 and 1.26.2 were malicious and removed from PyPI</td>
          <td>Confirmed</td>
          <td>Wiz, Chainguard, SecurityWeek; PyPI no longer lists them</td>
      </tr>
      <tr>
          <td>1.26.0 is the latest clean release</td>
          <td>Stated by Wiz and SecurityWeek</td>
          <td>Chainguard (writing earlier) did not name a safe version and advised suspicion of any release on Nov 17 or 18</td>
      </tr>
      <tr>
          <td>Microsoft is investigating</td>
          <td>Reported</td>
          <td>SecurityWeek, citing a maintainer&rsquo;s GitHub post; we found no Microsoft blog or advisory</td>
      </tr>
      <tr>
          <td>How publishing access was obtained</td>
          <td><strong>Not confirmed</strong></td>
          <td>Wiz: the attacker &ldquo;appears to have compromised a client certificate and private key&rdquo; for Microsoft&rsquo;s ESRP signing and publishing service. SecurityWeek: &ldquo;The method used to publish the malicious packages remains unknown.&rdquo;</td>
      </tr>
      <tr>
          <td>A PyPI account created on the day of the attack was involved</td>
          <td><strong>Not confirmed</strong></td>
          <td>SafeDep, as reported by SecurityWeek, found a new maintainer account with publishing rights; its role in the uploads is unconfirmed</td>
      </tr>
      <tr>
          <td>Victim data was stolen</td>
          <td>Partly reported</td>
          <td>Wiz says the webhook.site endpoint was on a free tier that caps at 50 requests, and it identified three unique successful exfiltrations among them</td>
      </tr>
      <tr>
          <td>Who is behind it</td>
          <td><strong>Not confirmed</strong></td>
          <td>No attribution published</td>
      </tr>
  </tbody>
</table>
<p>Wiz also mentions an earlier, separate compromise of a Microsoft package (durabletask) and says the two incidents have not been linked.</p>
<h2 id="how-the-stealer-ran">How the stealer ran</h2>
<p>Both wheels added a <code>.pth</code> file to <code>site-packages</code>: <code>agent_framework_meta.pth</code> in 1.26.1 and <code>af_runtime_diag.pth</code> in 1.26.2. Python&rsquo;s <a href="https://docs.python.org/3/library/site.html"><code>site</code> module documentation</a> explains why that is enough: in a <code>.pth</code> file, &ldquo;lines starting with <code>import</code> (followed by space or tab) are executed with <code>exec()</code>.&rdquo; The <code>site</code> module processes these files at every interpreter start, so the payload ran on any <code>python</code> invocation in that environment, including test runs, linters and unrelated scripts. Python 3.15 deprecates <code>import</code> lines in <code>.pth</code> files, but they still execute during the deprecation period.</p>
<p>Per Wiz&rsquo;s analysis:</p>
<ul>
<li><strong>1.26.1</strong> loaded <code>agent_framework_meta</code>, which called <code>_telemetry._start()</code>. It detached into a separate session or daemon thread and set <code>_AF_TEL=1</code> so it would not run twice.</li>
<li><strong>1.26.2</strong> loaded a new <code>af_runtime_diag.py</code> and restored the package&rsquo;s <code>__init__.py</code> to the clean 1.26.0 version. Hunting only for 1.26.1&rsquo;s file names misses it.</li>
</ul>
<p>What it collected, as described by Wiz and Chainguard:</p>
<ul>
<li>Environment variables whose names contain words such as TOKEN, SECRET, KEY, AZURE, AWS, GITHUB and NPM, plus secrets read from other processes&rsquo; <code>/proc/&lt;pid&gt;/environ</code>.</li>
<li>Credential files including Azure CLI configuration, <code>.npmrc</code>, <code>pip.conf</code>, <code>.netrc</code>, <code>.git-credentials</code>, Docker <code>config.json</code> and Hugging Face tokens.</li>
<li>Azure DevOps OIDC tokens and Azure managed-identity tokens. Wiz counts 12 Azure resources; Chainguard says nine resource types.</li>
<li>In 1.26.2, GitHub Actions OIDC tokens for <code>api://AzureADTokenExchange</code> and <code>https://management.azure.com/</code>.</li>
</ul>
<p>The data went out in one HTTPS POST to a webhook.site URL. Wiz notes that stolen Azure tokens may stay valid for up to 24 hours.</p>
<h2 id="indicators-of-compromise">Indicators of compromise</h2>
<p>From Wiz&rsquo;s write-up:</p>
<table>
  <thead>
      <tr>
          <th>Type</th>
          <th>Value</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td>Files</td>
          <td><code>agent_framework_meta.pth</code>, <code>agent_framework_meta/_telemetry.py</code>, <code>af_runtime_diag.pth</code>, <code>af_runtime_diag.py</code></td>
      </tr>
      <tr>
          <td>Environment</td>
          <td><code>_AF_TEL=1</code></td>
      </tr>
      <tr>
          <td>Network</td>
          <td><code>webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb</code>; User-Agents <code>af-runtime-diagnostics/1</code> (1.26.1) and <code>runtime-diagnostics/1</code> (1.26.2)</td>
      </tr>
      <tr>
          <td>1.26.1 wheel SHA-256</td>
          <td><code>d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110</code></td>
      </tr>
      <tr>
          <td>1.26.2 wheel SHA-256</td>
          <td><code>3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9</code></td>
      </tr>
      <tr>
          <td><code>af_runtime_diag.pth</code> SHA-256</td>
          <td><code>325e7e9f0ec94e41ad1a1b7a3f18a6b8b5bffece2e2b97b8d3369eb967452f93</code></td>
      </tr>
      <tr>
          <td><code>af_runtime_diag.py</code> SHA-256</td>
          <td><code>54d885b7d7161c9ed96489c66364c4bfbf635929166843bba10d97d46b5a2e1f</code></td>
      </tr>
  </tbody>
</table>
<p>Chainguard advises blocking the full webhook.site path rather than the whole domain if your teams use webhook.site for legitimate testing.</p>
<h2 id="what-to-check-in-your-environment">What to check in your environment</h2>
<p><strong>1. Find installs without starting Python.</strong> Running <code>pip</code> or any Python tool in an affected environment starts the interpreter and can run the payload. Search the filesystem with shell tools instead. These commands use GNU <code>find</code> and <code>grep</code>; we tested them against a mock directory tree containing the IOC file names, not against the real malware.</p>
<div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-bash" data-lang="bash"><span style="display:flex;"><span><span style="color:#75715e"># Files the two malicious wheels drop (search every filesystem you care about)</span>
</span></span><span style="display:flex;"><span>find / -xdev <span style="color:#ae81ff">\(</span> -name <span style="color:#e6db74">&#39;agent_framework_meta.pth&#39;</span> -o -name <span style="color:#e6db74">&#39;af_runtime_diag.pth&#39;</span> <span style="color:#ae81ff">\
</span></span></span><span style="display:flex;"><span><span style="color:#ae81ff"></span>  -o -name <span style="color:#e6db74">&#39;af_runtime_diag.py&#39;</span> -o -type d -name <span style="color:#e6db74">&#39;agent_framework_meta&#39;</span> <span style="color:#ae81ff">\
</span></span></span><span style="display:flex;"><span><span style="color:#ae81ff"></span>  -o -type d -name <span style="color:#e6db74">&#39;agent_framework-1.26.[12].dist-info&#39;</span> <span style="color:#ae81ff">\)</span> -print 2&gt;/dev/null
</span></span><span style="display:flex;"><span>
</span></span><span style="display:flex;"><span><span style="color:#75715e"># Pins and lockfile entries for the bad versions in a source tree</span>
</span></span><span style="display:flex;"><span>grep -rnE <span style="color:#e6db74">&#39;agent[-_]framework[ =&lt;&gt;~!&#34;]*1\.26\.[12]\b&#39;</span> <span style="color:#ae81ff">\
</span></span></span><span style="display:flex;"><span><span style="color:#ae81ff"></span>  --include<span style="color:#f92672">=</span><span style="color:#e6db74">&#39;*.txt&#39;</span> --include<span style="color:#f92672">=</span><span style="color:#e6db74">&#39;*.lock&#39;</span> --include<span style="color:#f92672">=</span><span style="color:#e6db74">&#39;*.toml&#39;</span> --include<span style="color:#f92672">=</span><span style="color:#e6db74">&#39;*.cfg&#39;</span> .
</span></span><span style="display:flex;"><span>grep -rn -A1 <span style="color:#e6db74">&#39;name = &#34;agent-framework&#34;&#39;</span> --include<span style="color:#f92672">=</span><span style="color:#e6db74">&#39;*.lock&#39;</span> . | grep -E <span style="color:#e6db74">&#39;version = &#34;1\.26\.[12]&#34;&#39;</span>
</span></span></code></pre></div><p>Also check places a filesystem search will not cover: CI job logs from November 17-18 (look for <code>agent-framework-1.26.1</code> or <code>1.26.2</code> in resolver output), container image layers built in that window, and internal package mirrors that may have cached the wheels.</p>
<p><strong>2. Check egress.</strong> Search proxy and DNS logs for the webhook.site path above. Wiz reads an HTTP 200 as data accepted and 410 or 429 as rejected, but treat any attempt as a reason to rotate.</p>
<p><strong>3. Clean, then rotate.</strong> Remove the <code>.pth</code> and module files, then rebuild the environment from a pinned, known-good set (<code>agent-framework==1.26.0</code>). Rotate everything the environment could read: Azure CLI sessions, Azure DevOps PATs and service connections, npm, PyPI, NuGet, Docker and Hugging Face tokens, git credentials and cloud keys. Our guide to <a href="/posts/how-to-detect-and-rotate-leaked-api-keys-before-attackers-use-them/">detecting and rotating leaked API keys</a> covers the order of operations.</p>
<p><strong>4. Review federated identities.</strong> For GitHub Actions workflows that installed 1.26.2 with <code>id-token: write</code>, Wiz says to treat the OIDC tokens as exposed and review the Azure federated credentials that trust those workflows. Check Azure sign-in and activity logs for managed identities reachable from affected hosts.</p>
<h2 id="what-engineers-can-take-from-it">What engineers can take from it</h2>
<ul>
<li><strong>Pin and hash-lock dependencies</strong> in CI and image builds (<code>pip install --require-hashes</code>, or a lockfile with hashes). A floating <code>agent-framework</code> requirement pulled whatever was newest during the window.</li>
<li><strong>A delay on brand-new versions helps.</strong> Both releases were gone within about an hour. A mirror or resolver rule that waits before accepting fresh uploads would have skipped them. Chainguard&rsquo;s 42 to 48 second blocks show what a curated mirror can do.</li>
<li><strong>Keep OIDC scopes narrow.</strong> Grant <code>id-token: write</code> only to jobs that deploy, and bind cloud federated credentials to specific repositories, branches or environments. Our overview of <a href="/posts/software-supply-chain-attacks-where-modern-build-pipelines-break/">where build pipelines break</a> covers the wider pattern.</li>
<li><strong>Trusted publishers are not proof of safety.</strong> If Wiz&rsquo;s reading of the ESRP credentials is right, the releases came through Microsoft&rsquo;s own publishing path. Provenance tells you where a package was built; it does not tell you the build was intended. See <a href="/posts/slsa-build-levels-and-provenance-attestations-what-they-prove-and-what-they-dont/">what SLSA provenance does and does not prove</a>.</li>
</ul>
<h2 id="sources-and-verification">Sources and verification</h2>
<table>
  <thead>
      <tr>
          <th>Important claim</th>
          <th>Source</th>
          <th>Status</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td>Affected versions, timeline, ESRP credential reading, payload behaviour, IOCs, three exfiltrations</td>
          <td><a href="https://www.wiz.io/blog/malicious-version-of-microsofts-agent-framework-published-to-pypi">Wiz</a></td>
          <td>Qualified (researcher analysis)</td>
      </tr>
      <tr>
          <td>Second-level timeline, .pth mechanism, &ldquo;remove the payload before starting Python again&rdquo;</td>
          <td><a href="https://www.chainguard.dev/unchained/microsoft-agent-framework-compromised-on-pypi-november-2026">Chainguard</a></td>
          <td>Qualified (researcher analysis)</td>
      </tr>
      <tr>
          <td>Maintainer confirmation, ~190,000 monthly downloads, publishing method unknown, SafeDep account finding</td>
          <td><a href="https://www.securityweek.com/microsoft-agent-framework-releases-poisoned-with-credential-stealer/">SecurityWeek</a></td>
          <td>Qualified (press report)</td>
      </tr>
      <tr>
          <td>1.26.0 is the latest listed release; 1.26.1 and 1.26.2 absent</td>
          <td>PyPI JSON API for agent-framework, checked Nov 19, 2026</td>
          <td>Verified</td>
      </tr>
      <tr>
          <td><code>.pth</code> import lines are executed at startup</td>
          <td><a href="https://docs.python.org/3/library/site.html">Python <code>site</code> documentation</a></td>
          <td>Verified (verbatim)</td>
      </tr>
  </tbody>
</table>
]]></content:encoded></item></channel></rss>