sysnote
News

Wiz ties the agent-framework uploads to a stolen Microsoft ESRP certificate; Microsoft made the ONNX repos private

Two questions were open when we wrote about the malicious agent-framework 1.26.1 and 1.26.2 releases on PyPI (our report): how someone got to publish under Microsoft's own project, and why microsoft/onnxruntime, microsoft/agent-framework and several other repositories started returning 404 the same night. Since then Wiz has published a root cause, and Microsoft has said it made the repositories private itself. Neither says how the attacker got the access in the first place.

The upload path

In an update posted on November 18 at 14:00 UTC, Wiz researchers Rami McCarthy, Scott Piper and Benjamin Read write that the attacker "appears to have compromised a client certificate and private key for Microsoft ESRP (Engineering System Release Pipeline)", the internal service Microsoft uses to sign and publish the package. The first publish attempt came from AWS and was rejected by ESRP's IP filtering. The second came from Microsoft-owned IP space and went through. How the certificate was obtained, Wiz says, is not yet known.

Wiz could see this much because the attacker's tooling reported its progress to the same public webhook.site URL it used for the stolen data. The tooling calls itself Cairn: it staged the package in /tmp/cairn_pkg and sent a connectivity probe named CAIRN_ESRP_PROBE2. Cairn is an open-source framework in which AI agent workers search toward a goal by writing facts and planned next steps to a shared board; its README lists Claude Code, Codex and Pi as worker backends and penetration testing as its first use. Before publishing, the attacker ran the stealer on their own machine. That run captured an agent configuration that Wiz reads as the open-source Pi harness with a GLM-family model, labelled pi_glm53, behind a local proxy.

Cairn has turned up in intrusions before. Palo Alto Networks' Unit 42 described a campaign it tracks as CL-STA-1238, in which attackers ran a customized fork of Cairn against at least 19 organizations in August and September, on Zhipu GLM 5.2 and DeepSeek V4 models. The framework is public, and Wiz does not connect the two cases.

The webhook.site URL was on the free tier, which keeps only the first 100 requests to a URL. Wiz analysed the 50 that were stored: four came from the attacker's own machines, 29 from sandboxes, vendor scanners and tests, and the remaining 17 from three victims whose data was taken. Wiz does not say whether more requests reached the URL than the service kept.

The PyPI account

The earlier explanation pointed elsewhere. SafeDep and StepSecurity named vibiret, a PyPI account created on November 17 and listed as a maintainer of agent-framework, and StepSecurity reported the uploader user-agent RestSharp/106.13.0.0 on both wheels. Wiz's post does not mention the account. The PyPA advisory PYSEC-2026-4388 says both versions were uploaded "without authorization" and that "PyPI Security worked with Microsoft to review the project's accounts and complete remediation steps."

After that review the account is still there. SafeDep's November 18 update listed agent-framework and five related packages for it. When we opened the PyPI profile page for vibiret at 04:56 UTC on November 19, it listed 14 projects: those six, agent-framework-eval, agent-framework-workflow and an archived agent-framework-azure, plus semantic-kernel, autogen-core, autogen-ext, autogen-agentchat and pyautogen. None of the 14 has had a release since November 12. semantic-kernel was last released on November 5, the AutoGen packages in 2025.

From public pages we can't tell whether SafeDep's list was incomplete or the account got more projects on November 18. An account run by the attacker is unlikely to have kept its access through a review by PyPI and Microsoft, still less to be listed on Semantic Kernel and AutoGen afterwards, so a Microsoft-controlled account looks more likely to us. Nobody has said so. Until someone does, SafeDep's advice still applies: lock the whole dependency tree with hashes, because agent-framework 1.26.0 accepts any release below 2.0 of four sibling packages that vibiret can publish.

The repositories

Microsoft gave The Register a statement on the evening of November 18: "We have made some repositories private while we investigate malicious content. Our priority is to protect customers and the broader ecosystem. We are working to review and will take further action as needed." The statement names no repositories and does not say what the malicious content is.

At 04:55 UTC on November 19 the seven repositories from our first report still returned 404, as did microsoft/onnxruntime-inference-examples. microsoft/Olive, microsoft/semantic-kernel, microsoft/autogen and microsoft/durabletask-python were up. On PyPI, onnxruntime 1.32.0 is still the latest release, and 1.26.0 is again the newest agent-framework. Anything that clones ONNX Runtime or downloads its release archives from GitHub has now been broken for more than a day.

Microsoft has not said how the ESRP certificate was taken, whether it has been revoked, or what it found in the repositories it took private.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.