Attackers Hijack Microsoft AI SDK to Steal Cloud Credentials
Microsoft's official AI agent SDK was hijacked on PyPI to distribute malware that exfiltrated cloud tokens and secrets from developer workstations.
Attackers hijacked Microsoft's official AI agent SDK and Agent Framework Python packages on the Python Package Index (PyPI) on November 17, 2026. The threat actors released two malicious versions, 1.26.1 and 1.26.2, which remained live for less than an hour before PyPI quarantined them.
The malware used .pth startup hooks to execute automatically upon Python initialization, even if the framework was not imported. The payload, disguised as diagnostics, harvested environment secrets, cloud tokens, and CI/CD credentials from platforms including Azure, AWS, GitHub, OpenAI, Hugging Face, and Docker. These secrets were then exfiltrated to an attacker-controlled webhook.
Security firms including OX Security, Aikido Security, Endor Labs, SafeDep, and StepSecurity identified the breach, noting that the attackers likely hijacked the wheel file build process. Researchers from Wiz identified at least three distinct cases of successful data exfiltration and suggested the attackers may have compromised a client certificate and private key from Microsoft's Engineering System Release Pipeline to gain access.
Microsoft has not yet issued an official statement regarding the cause of the breach. Security experts warn that any unpinned installations or cached container images from the attack window remain compromised, advising affected developers to rebuild environments and rotate all reachable secrets.