Developers building AI agents with Microsoft's Agent Framework need to check their environments today. Two malicious versions of the agent-framework package, 1.26.1 and 1.26.2, were published to PyPI on November 17, just 33 minutes apart.

Both carried a credential stealer that ran every time Python started, through a .pth startup hook, so the package did not even need to be imported. It went after Azure and Azure DevOps tokens and GitHub Actions secrets, and the second version added theft of GitHub Actions OIDC tokens. Stolen data was sent to a webhook.site address.

The releases were pulled within about 45 minutes and are now quarantined. Version 1.26.0 is the latest clean release. Microsoft has acknowledged the incident and says it is investigating. Researchers found no matching code in Microsoft's main repository, and the malicious builds shipped without source distributions. How the attacker got publishing access is still unknown.

The package sees about 190,000 downloads a month, so exposure could be wide in CI pipelines that pull the latest version automatically.

What to do: pin agent-framework to 1.26.0, search site-packages for agent_framework_meta.pth or af_runtime_diag.pth, and rotate every secret reachable from any machine or CI runner that installed either bad version.

Photo: Unsplash