AI & Agents

Microsoft’s Agent Framework Package on PyPI Was Poisoned With a Credential Stealer

Leave a comment

Attackers published two malicious versions of Microsoft’s official Agent Framework Python package to PyPI on Nov. 17: 1.26.1 and 1.26.2, 33 minutes apart. It’s the toolkit for building and orchestrating AI agents, with roughly 190,000–200,000 downloads a month.

The nasty part: the stealer ran on every Python startup through an agent_framework_meta.pth file, even if you never imported the package. It kept the dependency on the real agent-framework-core 1.26.0, so everything still worked while it ran. Version 1.26.1 went after secrets in environment variables, config files and other processes (Azure, AWS, GitHub, Docker, Hugging Face, OpenAI and more), plus Azure managed identity and Azure DevOps OIDC tokens. Version 1.26.2 added GitHub Actions OIDC token grabs and runner directory searches. Socket, StepSecurity and Ox Security documented it, and PyPI has quarantined both versions. How the attacker got publish access isn’t known yet.

If either version touched a dev box or CI runner, assume every secret it could reach is burned. Rotate from a clean machine, rebuild the environment, and pin with hashes. Agent tooling sits right next to the keys, which makes it a great target.

God Bless, and Needs Garlic!

References & Further Reading

Subscribe (RSS) Follow AI & Agents via RSS

Leave a note

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.