Threat actors recently compromised Microsoft’s official AI agent building project, Agent Framework, in a targeted supply chain intrusion. Specifically, attackers published malicious versions 1.26.1 and 1.26.2 directly to the Python Package Index (PyPI). These poisoned releases bundled a credential stealer designed to harvest cloud keys, local environment secrets, and GitHub Actions tokens.
Fortunately, security teams detected and removed the rogue packages in under one hour. Currently, Microsoft maintainers are investigating how the intruders breached the publishing pipeline. Importantly, this Agent Framework PyPI attack impacted only PyPI distributions. Meanwhile, Microsoft’s open-source GitHub repositories and the .NET edition remained completely unaffected.
Incident Timeline and Potential Developer Exposure
According to Socket’s security report on the compromised Agent Framework package, the attack unfolded rapidly on November 17, 2026:
- 20:35 UTC: Attackers uploaded malicious version 1.26.1 containing the credential stealer to PyPI.
- 21:08 UTC: Threat actors pushed a second compromised release, version 1.26.2, to the registry.
- 21:30 UTC: Responders identified and pulled both malicious versions offline, prompting an immediate Microsoft investigation.
Normally, Agent Framework averages roughly 41,000 weekly downloads and 200,000 monthly downloads. However, neither Microsoft nor PyPI has disclosed exact download counts during the 55-minute exposure window. Consequently, any developer who installed or updated the library during that timeframe must audit their systems immediately.
Suspected Compromise of PyPI Publishing Credentials
Although Microsoft has not yet confirmed the initial entry vector, forensic evidence points toward abused publishing credentials. Specifically, attackers likely exploited an exposed PyPI upload token, a misconfigured Trusted Publisher rule, or a compromised maintainer account. Notably, neither 1.26.1 nor 1.26.2 had matching GitHub tags, release notes, or standard release manifests. Because legitimate releases always generate source archives and repository tags, the attackers clearly bypassed standard release automation.
Why Developers Must Not Run Standard Pip Uninstall
Crucially, Google cloud security subsidiary Wiz warned affected developers against running standard pip commands to remove the package. Invoking pip automatically triggers startup execution of files embedded inside the malicious release, running the credential stealer again. Instead, security teams must follow a strict manual remediation sequence:
- Activate the Kill Switch: Set the environment variable
_AF_TEL=1immediately and manually delete the malicious startup files. - Pin a Clean Release: Reinstall the verified package version by running
agent-framework==1.26.0. - Purge Package Caches: Remove versions 1.26.1 and 1.26.2 from local pip caches, uv caches, CI runners, and internal mirrors.
- Rebuild Environments: Recreate any Docker containers and Python virtual environments that pulled the compromised releases.
- Rotate Exposed Secrets: Revoke and regenerate all cloud credentials, repository tokens, and CI/CD secrets on affected hosts.
- Audit Cloud Logs: Inspect Entra, Azure, Azure DevOps, and GitHub Actions logs for unauthorized activity after November 17, 2026, at 20:35 UTC.
Ultimately, rapid credential rotation and cache cleanup remain essential to containing exposure from short-lived Python supply chain attacks.
Support Our Threat Intelligence
Find our threat intelligence and malware analysis helpful? Support our work today and unlock a 100% ad-free reading experience!