Microsoft's AI Agent Framework Compromised: Supply Chain Attack Steals Cloud Credentials
Threat actors successfully compromised Microsoft's `agent-framework` package on the Python Package Index (PyPI), distributing malicious versions (1.26.1 and 1.26.2) that function as credential stealers. These versions were published on November 17, 2026, and were designed to execute a credential-stealing routine every time Python started, regardless of whether the package was explicitly imported. The malware targeted environment variables, credential files (including those for Azure CLI, npm, pip, Docker, Hugging Face, and Azure DevOps), and even requested fresh Azure DevOps OIDC and Azure managed-identity tokens. All collected data was then exfiltrated to an attacker-controlled webhook. Both malicious versions have since been quarantined from PyPI.
This incident is significant because it directly targets the software supply chain for AI development, specifically impacting users of Microsoft's Agent Framework, an open-source tool for building and orchestrating AI agents. The automatic execution mechanism upon Python startup makes this attack particularly insidious, as it could compromise development and CI/CD environments without immediate user awareness. For DevOps and cloud engineers, this means that any system that installed these specific versions during the brief window they were available is at risk. The compromise of a widely used, official package from a major vendor like Microsoft erodes trust in public package repositories and emphasizes the need for rigorous validation of all dependencies.
This attack fits into a broader, well-established trend of software supply chain compromises, which have seen a significant uptick in recent years. Attackers are increasingly targeting package managers and open-source repositories as a high-leverage way to distribute malware, impacting a wide array of downstream users. The rise of AI development further exacerbates this trend, as new AI-specific libraries and frameworks become attractive targets. Previous incidents, such as the compromise of Microsoft's `durabletask` package in May 2026, indicate a persistent focus by attackers on the Microsoft Python ecosystem. The shift towards AI-driven development means that security practices must evolve to cover not just traditional application code but also the entire AI development lifecycle, including model dependencies and agent frameworks.
In practice, practitioners must immediately audit their environments to identify any installations of `agent-framework` versions 1.26.1 or 1.26.2. If found, affected systems should be isolated, and all credentials that could have been exposed (including environment variables, cloud credentials, and API tokens) must be rotated from a clean system. Pinning package versions to known safe releases (e.g., 1.26.0) and implementing robust software supply chain security tools that perform real-time scanning and integrity checks on downloaded packages are crucial. Furthermore, organizations should enhance their monitoring of CI/CD pipelines and developer workstations for anomalous outbound connections or credential usage. This incident serves as a stark reminder that even official packages can be compromised, necessitating a defense-in-depth approach to supply chain security.
Read original source