Intelligence Report

Attackers push credential-stealing releases of Microsoft's AI agent toolkit to PyPI

Two tainted Agent Framework versions went after cloud, GitHub and AI service secrets before PyPI quarantined both within minutes.

INSIGHTEX Editorial Desk

2 min read

Attackers push credential-stealing releases of Microsoft's AI agent toolkit to PyPIIntelligence Report

What happened

Attackers gained the ability to publish to Microsoft's official Agent Framework project on PyPI and pushed two tainted releases, 1.26.1 and 1.26.2, on November 17, 2026. Both carried code built to harvest cloud credentials and other secrets from developer machines, BleepingComputer reported. Security firms Socket, Step Security and Ox Security each identified and documented the attack.

The 1.26.1 release pulled secrets out of environment variables, config files stored locally and, when it had sufficient permissions, other processes running on the machine. It targeted credentials for services including Azure, AWS, GitHub, Docker, Hugging Face and OpenAI. According to Socket and StepSecurity, it used a .pth file to launch at Python startup, so it ran even if the package was never imported. Meanwhile, because it still relied on the genuine agent-framework-core 1.26.0 package, the framework kept functioning as expected.

Version 1.26.2 moved the malicious code into new files. It also added requests for GitHub Actions OIDC tokens and searches of GitHub runner directories. PyPI quarantined the first release, but the second appeared about 2.5 minutes later and was itself quarantined roughly eight minutes after that, which suggests the attacker still held publishing rights. How the attacker obtained access remains unknown, and Microsoft had not answered questions about the point of breach.

Context

Microsoft Agent Framework is an open-source Microsoft toolkit that developers use to create AI agents and coordinate them within workflows, and it draws roughly 200,000 monthly downloads on PyPI. Affected developers were advised to isolate their environments, avoid running Python or pip until those systems are secured, and rebuild from the clean 1.26.0 release. They should also rotate any exposed credentials and check egress logs for traffic to the attacker's webhook.

Docker profile

GitHub profile

SourceBleepingComputer

The briefing

The week in tech, in one email.

The stories worth your time, every week. Unsubscribe anytime.