Who Can Help Address The Microsoft AI Agent Framework PyPI Attack?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Who Can Help Address The Microsoft AI Agent Framework PyPI Attack? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Two versions of Microsoft’s Python package agent-framework, 1.26.1 and 1.26.2, contained credential-stealing code and were quarantined by PyPI on November 17, 2026. Security advisories say affected environments should be treated as exposed, rebuilt from a clean base and have accessible credentials rotated; the number of affected users is not known.

Two releases of Microsoft’s agent-framework Python package carried credential-stealing code before PyPI quarantined them on November 17, 2026, according to security advisories and researchers who analyzed the packages. The affected versions, 1.26.1 and 1.26.2, could expose cloud, developer and AI-service credentials from environments that installed them, making the incident a concern for developers, organizations and CI build systems.

The GitHub Advisory Database says the malicious code was absent from the preceding clean release, 1.26.0, and rates the incident critical. PyPI records and researcher timelines cited in the source report place publication of version 1.26.1 at 20:35 UTC. StepSecurity says PyPI quarantined it at 21:06 UTC. A second release, 1.26.2, appeared at 21:08 UTC and was quarantined at 21:16 UTC. The report says the first release was available for about 31 minutes.

The advisory says a Python start-up hook caused the code to run when Python launched in an affected environment, even if a user did not import the framework itself. The framework continued to work normally, so a successful installation or ordinary test would not necessarily reveal the compromise. Researchers describe the code as collecting accessible credentials and tokens from developer and cloud environments and sending them to an attacker-controlled collection address.

The advisory recommends treating any environment that installed either affected version as exposed. That includes laptops, build servers, CI pipelines and container images, including those updated automatically. Its response guidance is to rebuild from a clean base using version 1.26.0 and rotate every credential the environment could access, including CI and GitHub Actions secrets. The source report says both malicious versions have been removed from PyPI and identifies 1.26.0 as the current release there.

At a glance
reportWhen: Incident occurred November 17, 2026; bo…
The developmentSecurity researchers reported that two malicious releases of Microsoft’s agent-framework package briefly appeared on PyPI before being quarantined.

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 Systems Face Elevated Risk

The incident matters beyond users who deliberately installed an AI framework. A compromised development environment can hold credentials for cloud accounts, code repositories, package registries and AI services. If those credentials are stolen, an attacker may be able to access resources beyond the computer where the package was installed. The exact reach depends on which credentials were present and what permissions they had.

Build systems warrant particular attention because they can contain deployment secrets and credentials used across projects. The GitHub Advisory Database specifically advises rotating reachable credentials, including CI and GitHub Actions secrets. Researchers quoted in the source report say the malicious releases also sought short-lived tokens, which can still be useful if captured and exfiltrated promptly. This is a risk assessment, not confirmation that any particular account or service was accessed.

The brief availability period does not by itself establish the number of installations. A package may be pulled by automated workflows, and the framework’s continued operation could make the malicious behavior less obvious. Organizations therefore need to check package records and build histories rather than assume that a functioning application means its environment was unaffected.

Products Worth Considering

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

As an affiliate, we earn on qualifying purchases.

How the Two Releases Appeared

Microsoft’s Agent Framework is an open-source Python toolkit for building AI agents and coordinating them in workflows. The PyPI package agent-framework is an umbrella package that brings in framework components and integrations. The source report says Microsoft’s installation instructions point users to this package, distinguishing it from an unrelated lookalike.

The clean version 1.26.0 was published on November 12, according to the report. Version 1.26.1 appeared on November 17 as a wheel-only release, unlike the earlier releases examined by Endor Labs, which included source code. After PyPI quarantined it, the attacker uploaded version 1.26.2 roughly two minutes later, according to StepSecurity. StepSecurity and JFrog both warn that quarantining a release does not itself revoke the publisher’s compromised publishing access; the second upload illustrates why that distinction matters.

The OpenSSF malware database listed the incident as MAL-2026-20101, and the GitHub Advisory Database published GHSA-764m-3ppc-5m73. The source report says Socket, StepSecurity, Wiz, JFrog and Endor Labs analyzed the releases. Wiz said it promptly reported its findings to Microsoft on November 17. These are attributed accounts from researchers and databases, rather than a public incident report from Microsoft.

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

— GitHub Advisory Database

Victim Count and Credential Use

The source material does not give a confirmed number of affected installations or victims, nor does it establish how many credentials were successfully collected or used. The available incident description explains what the malware sought, but that does not prove that every type of credential was present or obtained in each environment.

It is also unclear from the source report which organizations installed the releases, whether any stolen credentials were used to access accounts or systems, and whether Microsoft has completed its investigation. The report says researchers found no remote-code-loading second stage in the code they inspected and describes the malware as not persisting beyond the installed package. Those findings do not rule out exposure or misuse of credentials stolen while it was present.

Microsoft’s reported statement refers to repositories being made private but does not specify which repositories, what publishing credential was compromised, or what remediation has been completed. The source report says it found no Microsoft security advisory, so readers should distinguish the researchers’ findings and database guidance from a detailed public account by Microsoft.

Audit, Rebuild, and Rotate Credentials

Teams should check dependency files, package caches, installation logs, CI job records and container build histories for versions 1.26.1 and 1.26.2. Any environment that installed either release should be treated as exposed under the advisory’s guidance, even if the framework appeared to run normally. Rebuild affected systems from a clean base with version 1.26.0 rather than relying only on removing the package.

Organizations should rotate credentials that were accessible from affected environments, including cloud keys, repository and CI tokens, AI-service credentials and package-registry credentials. They should also review relevant account and service logs for activity that cannot be explained by normal use. The source material does not provide a complete incident-response checklist, so teams should consult the linked security advisories and their own security responders for technical details.

Further clarity depends on updates from Microsoft, PyPI and the researchers, including whether the publisher’s access was secured, whether investigators identify affected users, and whether there is evidence of credential misuse. Until those details are published, the confirmed release window and advisory guidance are the clearest basis for response.

Key Questions

Which package versions were affected?

The reported malicious releases were agent-framework 1.26.1 and 1.26.2. The GitHub Advisory Database identifies 1.26.0 as the last clean release before them.

How could the malicious code run?

The advisory says the releases added a Python start-up hook. It could run when Python started in an affected environment, even if the user did not import agent_framework.

What should a team do if it installed an affected version?

The advisory’s guidance is to treat the environment as exposed, rebuild it from a clean base using version 1.26.0, and rotate every credential it could access, including CI and GitHub Actions secrets.

Is there a confirmed victim count?

No confirmed victim or installation count is provided in the source report. It also does not establish how many credentials were successfully collected or whether any stolen credentials were used.

Has Microsoft published a full public advisory?

The source report says Microsoft told The Register it had made some repositories private while investigating malicious content, but that the report found no Microsoft security advisory at publication time.

Source: ThorstenMeyerAI.com

You May Also Like

Real-SWE: Benchmarking AI Models On Private, Real-world, Enterprise Codebases

Emerging trend involves evaluating AI models on real, private enterprise codebases to assess practical performance, sparking increased industry interest.

Top Strategies For Growing Online Storage To Support Massive AI Chat Platforms

OpenAI details how it expanded its storage infrastructure to support over 1 billion ChatGPT users, focusing on architecture, capacity, and operational lessons.

13 Ways ChatGPT Shaped Consultants’ Data Science Skills

A randomized BCG trial found ChatGPT raised consultants’ data science scores, but a follow-up test found no measurable knowledge gain without the tool.

When It Came To Cal AI’s Viral Success, A Small Army Of Influencers Paid Off

Founder Zach Yadegari said paid influencer posts helped Cal AI grow to more than $30 million in revenue before its MyFitnessPal acquisition.