Two Microsoft agent-framework releases poisoned to steal secrets at Python startup
Two malicious agent-framework releases appeared on PyPI 33 minutes apart before being quarantined. Their code could run at Python startup without an application importing the framework. The earlier 1.26.0 lacks the added code, but reverting versions does not revoke exposed credentials.
A Microsoft agent-development package could steal credentials before an application deliberately used it. GitHub’s advisory identifies two malicious releases of the legitimate PyPI package agent-framework: 1.26.1 and 1.26.2. This was not a misspelled imitation package. The dangerous content appeared under a name developers already trusted.
Agent Framework supports AI-agent and multi-agent applications. Package managers bring that framework and its dependencies into developer machines and automated build environments. Source code and a distributed package are different objects: a design drawing is not the finished product. Wiz reports that the malicious code was absent from the project’s main branch. Reviewing the drawing does not automatically authenticate every delivered batch.
Wiz places the first malicious upload at 20:35 UTC on November 17 and the second at 21:08, 33 minutes later. By approximately 21:30, neither was available from PyPI. That window corresponds to 04:35–05:30 Beijing time on November 18. PyPI’s release record marks both versions as quarantined; the current page presents 1.26.0, previously published on November 12.
The trap was attached to Python startup. A startup hook could run the collector when the interpreter started, without the application importing the framework. Think of equipment that switches on with the warehouse lights, rather than waiting for someone to operate it. The advisory describes collection of secrets, cloud credentials and automation tokens. The second release added collection targeting GitHub Actions identity tokens.
Quarantining releases restricts future distribution; it does not clean environments that already installed them. The advisory identifies these two malicious versions, while 1.26.0 predates the incident and lacks their added code. It is not a newly issued repair. Teams need to establish which machines or builds obtained the affected releases and what authority those environments held. Reverting a dependency cannot invalidate credentials that may already have escaped.
iHolyshit review
The more work agents can perform, the more valuable the authority around their development environments becomes. The lesson is not another behavioural instruction for an AI. Publishing, installation and credential handling belong to one security chain. A familiar package name is not a guarantee about a particular release.
Share this article
What do you think of this story?
Anonymous feedback, up to 5 reactions per article.
Reader comments
Comment without signing in, or sign in to display your member name.
Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.
No comments yet. Be the first to share.