- Get link
- X
- Other Apps
Featured Post
By Mesum Mukhtar
LearnDailyAI
-
- Get link
- X
- Other Apps
Two malicious releases of Microsoft's agent framework package appeared on PyPI on 17 November 2026. Versions 1.26.1 and 1.26.2 shipped with a credential stealer that ran every time Python started, even if the developer never imported the framework. Both versions were pulled within about an hour.
Agent framework is the meta package for Microsoft Agent Framework, an open source framework for building AI agents using Python. The package sees close to 200,000 downloads a month, so a poisoned release in that slot is a serious supply chain incident.
What happened
The two versions were published 33 minutes apart on 17 November. Version 1.26.1 appeared at 20:35 UTC and version 1.26.2 at 21:08 UTC, and both were gone from PyPI by about 21:30 UTC.
The attack hid in a Python startup hook. Both releases dropped a .pth file into the site packages folder, and Python runs executable lines in those files every time the interpreter starts. Version 1.26.1 used agent_framework_meta.pth and version 1.26.2 used af_runtime_diag.pth. The framework kept working, so developers saw nothing wrong.
Once running, the stealer collected a wide haul. It gathered environment variables that look like credentials, credential files for Azure CLI, npm, pip, netrc, Git, Docker and Hugging Face, plus Azure DevOps tokens and Azure managed identity tokens. Version 1.26.2 also requested GitHub Actions tokens. The haul went out in one HTTPS POST to a webhook.site address.
| Version | Published 17 Nov 2026 | Startup hook |
|---|---|---|
| 1.26.1 | 20:35 UTC | agent_framework_meta.pth |
| 1.26.2 | 21:08 UTC | af_runtime_diag.pth |
| 1.26.0 | Clean release | None |
Why it matters
The victim count looks small, but the method is the story. Wiz analysed the 50 requests stored at the attacker's webhook endpoint and found 17 of them represented only 3 unique successful thefts. The rest came from the attacker's own machines and from sandboxes and scanners.
The bigger shock is how the attacker published at all. Wiz assesses that the attacker appears to have used a client certificate and private key for Microsoft's Engineering System Release Pipeline. Address filtering blocked a first attempt from Amazon Web Services, but a second attempt from Microsoft owned address space succeeded. Microsoft has not published a root cause. This is the second Microsoft related PyPI compromise this year.
What to watch
If you use Python, check your installs now. Pin the package to 1.26.0, treat any environment that installed 1.26.1 or 1.26.2 as exposed, and rotate every credential that environment could reach. Check for exposure between 8:35 and roughly 9:20 p.m. UTC on 17 November 2026. Rebuild from a clean environment.
Watch for Microsoft's root cause report. Until Microsoft explains how its release pipeline credential was obtained, every team that publishes packages should review its own signing and publishing setup. The lessons are clear: use trusted publishing with build provenance, pin exact versions with cryptographic hashes, and delay adoption of brand new releases. Compare it with the agent tools in our earlier post on GPT Live 1.1 in the OpenAI API.
Watch the official video
Microsoft Reactor shows how to build practical AI agents with the Microsoft Agent Framework.
Watch "Getting Started with Microsoft Agent Framework" on YouTube
Sources
Wiz: Malicious version of Microsoft's Agent Framework on PyPI
The Hacker News: Attackers compromise Microsoft's Agent Framework package
TechRepublic: Microsoft Agent Framework attack
Follow Learn Daily AI on Instagram at @learningdailyai for one new AI story every day.
Comments
Post a Comment