Microsoft’s PyPI AI Agent Framework Incident: Who Needs To Act?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Microsoft’s PyPI AI Agent Framework Incident: Who Needs To Act? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

PyPI versions 1.26.1 and 1.26.2 of Microsoft’s agent-framework package contained credential-stealing code and were quarantined on November 17, according to security advisories and researchers. Anyone whose machines, pipelines or images installed either release should treat accessible credentials as exposed, rebuild from a clean version and rotate those credentials.

Two releases of Microsoft’s Agent Framework Python package on the public PyPI registry contained code that stole credentials, according to the GitHub Advisory Database and security researchers. The versions, 1.26.1 and 1.26.2, were quarantined on November 17, but developers and organizations that installed either release should treat the environment and credentials it could access as exposed.

The affected package, agent-framework, is the official Microsoft-published umbrella package for the company’s open-source toolkit for building AI agents in Python. The GitHub advisory says the malicious code was absent from the preceding clean release, 1.26.0. The bad versions were published at 20:35 and 21:08 UTC on November 17. PyPI’s current version is 1.26.0, released November 12, according to the source material.

Researchers’ timelines indicate PyPI quarantined version 1.26.1 at 21:06 UTC. An attacker uploaded 1.26.2 about two minutes later; that release was quarantined at 21:16 UTC, according to StepSecurity. Quarantine prevents installation through PyPI, but did not immediately prevent another upload. StepSecurity and JFrog said the first quarantine did not revoke the publishing access used to upload the package.

The advisory says the malicious code ran when Python started in an affected environment, even if the user did not import the framework. It sought credentials and tokens available to developer machines and build systems, including cloud, source-control and AI-service credentials. The advisory recommends treating any environment that installed either version as exposed, rebuilding it from a clean base using 1.26.0, and rotating every credential it could reach, including CI and GitHub Actions secrets.

At a glance
reportWhen: Incident occurred November 17, 2026; ad…
The developmentTwo malicious releases of Microsoft’s Python Agent Framework package were uploaded to PyPI and removed after a short exposure window on November 17.

agent-framework on PyPI · UTC

Two poisoned releases, one evening

The official package of Microsoft’s Agent Framework, step by step. Blue: what happened. Green: the clean version.

November 12 to 18, 2026

  1. Clean 1.26.0 publishedStill the current version on PyPI
  2. Malicious 1.26.1 appearsFramework works normally, a credential stealer runs whenever Python starts
  3. PyPI quarantines 1.26.1After about 31 minutes
  4. Attacker uploads 1.26.2Renamed files, also targets GitHub Actions tokens
  5. PyPI quarantines 1.26.2Eight minutes later
  6. Malware databases list itMAL-2026-20101 (OpenSSF), GHSA-764m-3ppc-5m73 (GitHub)
Quarantine is not revocation. The second release came 2 minutes 23 seconds after the first was pulled: the attacker still had publishing access. Copies already in caches, mirrors and images stay there.

Sources: GitHub Advisory Database GHSA-764m-3ppc-5m73, Nov 18, 2026; OSV MAL-2026-20101, Nov 17, 2026; PyPI project data, checked Nov 19, 2026; StepSecurity analysis (quarantine times).

Graphic: Powered by Thorsten Meyer AI

Why Build Pipelines Face Exposure

The incident matters because Python packages are often installed automatically in laptops, continuous-integration jobs and container builds. A package can therefore reach secrets that are not visible to an ordinary end user, such as deployment credentials or tokens used to publish software. The framework continued to function, according to the advisory, so a successful install or apparently normal workflow would not by itself show that an environment was clean.

Credential rotation is a separate task from removing the affected package. If a token was copied while it was available to the malicious code, uninstalling or rebuilding alone cannot make that old credential safe. Organizations need to identify where the affected versions were installed, determine what secrets those environments could access, and replace those credentials. The advisory and researchers describe the behavior; the available source material does not establish how many credentials were actually used or misused.

The short interval between uploads also shows why package quarantine and publisher-access controls matter together. JFrog and StepSecurity said blocking one version did not itself revoke the compromised publishing credential. That account does not establish how the attacker obtained publishing access, but it explains why responders should verify that the package source and its release controls are safe before trusting later releases.

Products Worth Considering

DJI Mini 4K Drone Bundle with Strobe Light, Landing Gear, and Propeller Holder Strap

DJI Mini 4K Drone Bundle with Strobe Light, Landing Gear, and Propeller Holder Strap

As an affiliate, we earn on qualifying purchases.

How the Two Releases Appeared

Microsoft’s Agent Framework is an open-source toolkit for building AI agents and coordinating workflows. The PyPI package agent-framework is a thin umbrella package that brings in the framework’s core code and integrations. The source material says Microsoft’s installation instructions point users to this official package, rather than to an unrelated lookalike.

According to the advisory and researcher timelines, Microsoft’s clean 1.26.0 release was published November 12. Version 1.26.1 appeared on November 17 as a wheel-only release, the advisory says. StepSecurity records its quarantine at 21:06 UTC, followed by the 1.26.2 upload at 21:08 and quarantine at 21:16. The GitHub Advisory Database rated the case critical; the OpenSSF malware database listed it as MAL-2026-20101. Security firms including Socket, StepSecurity, Wiz, JFrog and Endor Labs analyzed the releases.

Wiz said it promptly reported its findings to Microsoft on November 17. The source material also quotes Microsoft telling The Register that it had made some repositories private while investigating malicious content. It says no Microsoft security advisory had been found at the time of publication. These statements describe the response reported in the source and do not establish the full investigation’s conclusions.

“The credential stealer is not present in the last clean release, 1.26.0.”

— GitHub Advisory Database

Victim Count and Access Still Unknown

The source material gives no confirmed victim count and does not say how many downloads occurred during the affected window. Wiz describes the package as receiving almost 200,000 downloads a month, but that is a monthly figure, not a count of installations of the two malicious versions. The number of affected environments, and the number of credentials accessed or used, remain unclear.

The material does not identify how the attacker obtained publishing access, whether any stolen credentials were used for further activity, or whether Microsoft’s investigation has established a broader impact. It also reports no detailed Microsoft security advisory at the time covered by the source. Researchers’ findings about the inspected code should not be read as confirmation that every affected installation or downstream system has been identified.

What Affected Teams Should Do

Teams should check package records, lockfiles, build logs and container histories for versions 1.26.1 and 1.26.2, including installations made through automated dependency updates. Any environment that installed either version should be isolated as appropriate, rebuilt from a clean base pinned to 1.26.0, and reviewed for the credentials it could access. Following the advisory, rotate reachable credentials, including cloud, source-control, AI-service and CI secrets; review relevant access and audit logs for suspicious use.

Organizations should follow updates from Microsoft, PyPI, the GitHub Advisory Database and the security firms that analyzed the package. Further guidance may clarify the attacker’s access, the affected population and whether credentials were used after collection. Until then, the practical response is to treat potentially accessible secrets as exposed rather than assume that the short publishing window limited the impact.

Key Questions

Which package versions were affected?

The affected PyPI releases were agent-framework 1.26.1 and 1.26.2. The source material identifies 1.26.0, released November 12, as the preceding clean version and the current PyPI release at the time described.

Who should investigate?

Any individual or organization whose laptop, build server, CI pipeline or container image installed either affected version should investigate. Automated installs and updates may also have brought the releases into an environment.

Is uninstalling the package enough?

No. The advisory recommends rebuilding affected environments from a clean base and rotating every credential the environment could access. Removing the package does not invalidate a token that may already have been copied.

How many users were affected?

A confirmed number is not available in the source material. Wiz’s estimate of almost 200,000 downloads a month describes the package’s monthly volume, not downloads of the two malicious releases.

Source: ThorstenMeyerAI.com

You May Also Like

How To Pick An External GPU For Laptop Graphics In 2027: 10 Best Options

A buyer’s guide to 10 external GPU docks and enclosures, from ready-to-use RX 7600M XT systems to open OCuLink and M.2 designs.

Tailscale Database Corruption: SQLite WAL Issue

A brief flags a reported Tailscale database-corruption issue tied to SQLite WAL resets, but provides no technical advisory or incident details.

What We Get From Work—and What AI Might Take Away

An unreviewed survey of 8,234 adults finds many would leave work for guaranteed income, while a minority place lasting value on purpose and community.

Is ChatGPT Near A Fifth Of Google’s Search Volume? We Checked

OpenAI says ChatGPT handles search-like queries equal to one-fifth of Google’s volume, but the comparison is undefined and outside estimates vary.