Microsoft's official Python toolkit for building AI agents became a supply-chain weapon this week. Attackers published two compromised releases of the agent-framework package on PyPI — versions 1.26.1 and 1.26.2 — each laced with a credential stealer that executed automatically the moment a Python interpreter started, even if the developer never imported the package. The framework draws roughly 200,000 downloads per month, putting developer machines and CI/CD pipelines across the multi-cloud ecosystem in the blast radius.
The compromise was flagged on November 17 by supply-chain security firms Socket, StepSecurity and Ox Security. The first malicious release planted a file called agent_framework_meta.pth, which runs during Python startup and loads a hidden telemetry module that harvests environment variables, configuration files and process memory. Crucially, the package also pulled in the genuine agent-framework-core 1.26.0 release, so everything kept working normally while secrets flowed out in the background — a detail that made the tampering extremely hard to notice.
What got taken is a shopping list of cloud credentials: Azure, AWS, GitHub, Docker, Hugging Face and OpenAI keys, plus Azure managed-identity tokens and OIDC tokens from Azure DevOps. Version 1.26.2, uploaded after the first was quarantined, went further — it actively requested fresh GitHub Actions OIDC tokens and scanned GitHub runner directories, which makes short-lived CI tokens attractive even though they expire quickly. All collected data was base64-encoded and exfiltrated to a hard-coded webhook.site endpoint controlled by the attacker.
The response timeline shows how persistent the intruder was. PyPI quarantined version 1.26.1, and the attacker published the second malicious version roughly 2.5 minutes later — proof that publishing access had not been cut off. That release was quarantined in turn about eight minutes after appearing. The Register reported the incident appears to involve unauthorized access to an administrator account on Microsoft's GitHub organization; Microsoft had not responded to questions at publication time. Both the agent-framework repository and, separately, Microsoft's ONNX Runtime repository went offline, though no one has linked the ONNX outage to this campaign.
The target itself is telling. Agent frameworks are among the fastest-growing dependency categories in enterprise software, and they sit unusually close to privilege: the agents they build hold cloud credentials, connect to internal systems and run inside CI pipelines with permission to deploy code. A compromised framework package is therefore not just another library takeover — it is a pre-positioned foothold in the layer that orchestrates enterprise automation.
For teams that installed either version, the remediation is blunt. Security teams should treat every secret present in an affected environment as stolen: rebuild from the clean 1.26.0 release in a fresh environment, rotate Azure, GitHub, Docker and Hugging Face credentials, audit GitHub Actions and Azure DevOps token activity, and check egress logs for the known webhook URL to distinguish blocked requests from completed thefts. Deleting the package alone is not enough, because tokens already exfiltrated can be used the moment they are minted.
The episode lands the same week Microsoft pitched agentic AI as the future of work at its Ignite conference, and it underscores an uncomfortable asymmetry: the industry is racing to give agents real credentials and real authority, while the software supply chain that delivers the frameworks behind them still relies on trust in a publisher name. The lesson of 1.26.1 is that the official publisher account is no longer a sufficient safety guarantee.
Comments (0)
Log in to join the discussion
Log InNo comments yet