Executive Summary
On November 17, 2026, the official Microsoft Python package 'agent-framework' was compromised on PyPI. Malicious versions 1.26.1 and 1.26.2 were uploaded, incorporating a sophisticated credential stealer. The attack is notable for its persistence mechanism; it utilizes .pth startup hooks to execute the payload automatically whenever the Python interpreter starts, requiring no explicit import of the compromised library.
The malware targets a broad range of cloud and CI/CD secrets, including Azure managed identity tokens, GitHub Actions OIDC tokens, and environmental variables related to AWS, OpenAI, and Hugging Face. Despite PyPI quarantining the initial malicious release, the attacker successfully uploaded a second version with expanded capabilities minutes later, indicating a persistent breach of publishing credentials or flow.
This incident represents a critical risk to AI developers and CI/CD pipelines. Organizations using the Agent Framework must immediately audit their environments for the affected versions and treat any credentials accessible by those environments as compromised. Immediate rotation of Azure service principal secrets, PATs, and cloud tokens is highly recommended.