Two poisoned Microsoft Agent Framework releases hit PyPI
Two poisoned Microsoft Agent Framework releases reached PyPI on Nov. 17, exposing Azure, DevOps and GitHub credentials before both packages were removed and quarantined.
- By
- Grace Kim
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
By Grace Kim
Two poisoned releases of Microsoft’s official Python package for building AI agents reached PyPI on Tuesday, Nov. 17, remaining available for roughly 55 minutes and 22 minutes, according to Wiz and OX Security.
Wiz says version 1.26.1 was uploaded at 20:35 UTC on Nov. 17. Version 1.26.2, a repackaged copy of 1.26.1, followed at about 21:08 UTC. Both were unavailable when Wiz checked at 21:30 UTC. PyPI has quarantined both releases.
That first exposure window ran for nearly the length of a one-hour meeting. Version 1.26.2’s window lasted less than half as long.
Version 1.26.0, published on Nov. 12, is the most recent release known to be clean. The package receives about 190,000 downloads each month.
In a GitHub discussion, a Microsoft maintainer confirmed: “The compromised releases had been removed from PyPI,” and added that “an investigation is under way.” SecurityWeek reported the confirmation. Microsoft has not publicly explained how the attacker uploaded the packages or obtained the required publishing access.
What could the malicious code reach?
The researchers found that both releases executed their payload whenever a Python interpreter started in the affected environment, even when the application never imported Agent Framework. They still installed the genuine framework, so applications continued working normally.
The stealer gathered:
- Secrets from environment variables and other running processes.
- Credential files for cloud command-line tools, package registries and source control.
- Azure managed identity tokens and Azure DevOps pipeline tokens.
- GitHub Actions workflows, a target added in version 1.26.2.
The code sent captured material to an external web service. Version 1.26.1 lacked the GitHub Actions collection, while 1.26.2 extended the theft to those workflows.
How did the packages reach PyPI?
OX Security, Wiz, Endor Labs and SafeDep each analyzed the malicious code. The releases contained only built packages, with no matching source release. Wiz found no matching releases or malicious code in Microsoft’s source repository.
That absence suggests the uploads bypassed the normal release process, although it does not establish the full intrusion path. SafeDep found a PyPI maintainer account created on the day of the attack, with publishing rights for Agent Framework, but the account’s role remains unconfirmed.
Wiz said: “The root cause is not confirmed.” Microsoft has not publicly identified how the publishing access was obtained.
What should affected teams do?
Wiz and OX Security advise teams to search lockfiles, CI logs, virtual environments and container images for versions 1.26.1 and 1.26.2. Remove both versions, rebuild affected environments and pin Agent Framework to 1.26.0 with hash checking.
Because any Python process started in an affected environment ran the payload, researchers say teams must rotate every secret those hosts and build runners could reach. Those secrets include:
- Cloud sessions, pipeline tokens and package-registry tokens.
- Git credentials used by source-control systems.
Teams should also review Azure sign-in and activity logs for managed identities and federated credentials. Wiz says stolen tokens may remain valid for up to 24 hours.
For GitHub Actions workflows that installed 1.26.2 with OIDC write permissions, Wiz says teams should treat every affected token as exposed. The compromise check must extend beyond installed package records because the payload did not require an application import.
Why does the incident matter for agent builders?
The incident landed in a busy week for software supply-chain security. On the same day, OpenJDK and Oracle shipped fixes for the Java4Shell flaw. That parallel action matters because Agent Framework sits inside a wider chain linking package registries, developer workstations, CI runners, cloud identities and the services used to build AI agents.
The immediate issue is whether a standard package installation can execute unreviewed code before an application imports it. The investigation must still establish how the uploads occurred. Until then, teams deploying agents on Azure must treat package provenance and credential rotation as one production security boundary.
Source: mlq.ai
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,116 articles
Related articles
backdoored-claude-vs-code-extension-hit-up-to-5-000-developers
Backdoored Claude VS Code Extension Hit Up to 5,000 Developers
fake-ps5-emulators-on-github-deliver-password-stealing-malware
Fake PS5 Emulators on GitHub Deliver Password-Stealing Malware

