Skip to content

Microsoft’s agent framework was poisoned, and the loot was cloud identities

Microsoft’s AI agent framework shipped a credential stealer on PyPI, and it went straight for cloud identities. Passwords were too retro.

On 17 November, versions 1.26.1 and 1.26.2 of agent-framework, the meta-package for Microsoft’s open-source Agent Framework, arrived with a stealer attached. Aikido and The Hacker News put the releases 33 minutes apart; BleepingComputer says about two and a half. Even the crime scene can’t agree on the stopwatch. Both versions are now quarantined or removed.

Each wheel drops a .pth file that Python runs on every start, imported or not. It collects Entra managed-identity tokens, Azure DevOps OIDC tokens and assorted credential files, then ships them to a webhook.site address, because bespoke exfiltration infrastructure is so last year.

How the attacker got publishing access is unknown. Neither source reports a Microsoft statement.

Every “agent identity” vendor should now explain what happens when the agent’s own toolchain is the burglar. If these versions touched your estate, rotate the tokens; upgrading is the polite gesture, not the fix.

Sources: The Hacker News · BleepingComputer · Aikido · Socket

Topics:

happy_hippo

AI for IAM and IAM for AI