Microsoft Agent Framework. The Python package Agent Framework, Microsoft’s open-source toolkit for creating AI agents and coordinating them in workflows, has been compromised. Hackers have published two malicious versions, 1.26.1 and 1.26.2, designed to steal cloud credentials and other sensitive information from developer systems.

The compromised versions were released on November 17, 2026, and contained software capable of exfiltrating secrets from environment variables, local configuration files, and other running processes with their respective permissions. The malware targeted services such as Azure, AWS, GitHub, Docker, Hugging Face, and OpenAI.
Microsoft Agent Framework: why it matters
In particular, version 1.26.1 collected credentials from various sources, while 1.26.2 added OIDC token requests for GitHub Actions and searches in GitHub runner directories. Both versions exfiltrated the collected data to the same endpoint controlled by the attacker: an external web service.
Although the first malicious version (1.26.1) was quarantined by PyPI, the attacker managed to publish a second release (1.26.2) approximately 2.5 minutes later, which was also blocked within eight minutes. This suggests that the attacker maintained publishing access to Microsoft’s official project on PyPI.
Companies and developers who installed either of the two compromised versions should immediately isolate the affected environments, avoid running Python or pip until they are secured, and rebuild everything from a clean environment using version 1.26.0. It is also advisable to rotate potentially exposed credentials.
What changes and what are the effects
To assess the extent of the compromise, it is recommended to analyze egress logs for the specific webhook URL and distinguish between blocked requests and completed ones.
* Content created with the assistance of artificial intelligence systems.
Hardware Ready Ready to Bench?