Back to News
News · AI Agents & Security

Microsoft Agent Framework Releases Poisoned With Credential Stealer

Two tampered PyPI releases of Microsoft's agent-framework package quietly harvested secrets and cloud tokens.

Source: SecurityWeek · November 18, 2026

SecurityWeek reports that two malicious releases of Microsoft's agent-framework Python package, versions 1.26.1 and 1.26.2, appeared on PyPI about 33 minutes apart on November 17, alongside the legitimate 1.26.0 release. Researchers found the malware ran automatically whenever the Python interpreter started, even without the package being imported, and it searched for secrets in environment variables and configuration files and credentials from other running processes, while also attempting to obtain Azure managed identity and Azure DevOps tokens. The second release extended the theft to GitHub Actions, and PyPI has since quarantined both versions.

The affected package sees roughly 190,000 downloads a month, and how the malicious releases reached PyPI remains unclear. Wiz found no matching release or malicious code in Microsoft's own source repository, while SafeDep flagged a maintainer account created the same day as the attack, though its role is unconfirmed. Microsoft has acknowledged the incident and said an investigation is underway.

Why it matters: agent frameworks are fast becoming core enterprise AI infrastructure, and a single poisoned release in that supply chain can reach developer secrets, cloud tokens and CI/CD pipelines within minutes of publishing. Pin and verify every agent framework version before it reaches production, and rotate credentials quickly if one slips through.