Thursday, Nov 19 | --:--
Back to home

Poisoned Microsoft AI Agent SDK Releases Stole Cloud Secrets; Wiz Cites a Signing Cert

Versions 1.26.1 and 1.26.2 of agent-framework, Microsoft's Python package for building AI agents, were published to PyPI on November 17 with a credential stealer that runs whenever Python starts, collecting environment secrets, Azure tokens and, in the second release, GitHub Actions OIDC tokens. Both were removed within about an hour. Wiz says the attacker appears to have used a compromised client certificate for Microsoft's internal release pipeline, and that the tooling identified itself as Cairn, an open-source AI penetration-testing framework. Wiz counted three unique successful exfiltrations among the requests it could see. Version 1.26.0 is the last clean release.

5 min read Global View as Markdown
A handmade editorial collage on torn ivory paper in cobalt and vermilion: a plain brown paper-wrapped parcel bound with two parallel twine bands sits on a rough stone slab, and a thin clear tube from a seam at the parcel's side drips glowing cobalt droplets into a small blue bucket beside the slab; an official Microsoft logo is glued on a torn-paper clipping tucked under the stone slab's lower edge at bottom left, the full wordmark visible.

The package developers install to build AI agents became, for about an hour, a way to rob the machines that run them. Two releases of agent-framework, the Python package for Microsoft Agent Framework, were published to PyPI on November 17 carrying a credential stealer. What makes this one different from the run of poisoned AI libraries this autumn is how it was shipped: Wiz says the attacker appears to have published through Microsoft's own signing pipeline, using tooling that describes itself as an AI penetration-testing framework.

What happened#

Version 1.26.1 was uploaded at 20:35 UTC and 1.26.2 at about 21:08 UTC, and both were no longer available on PyPI by about 21:30 UTC, according to Wiz. GitHub's advisory, GHSA-764m-3ppc-5m73, lists both versions as malicious and names 1.26.0 as the last clean release. The package gets about 41,000 downloads a week, Wiz says; SecurityWeek put it at about 190,000 a month.

The stealer drops a .pth file into site-packages, which Python executes on every interpreter start, so it runs even if the application never imports the package. It collects secret-looking environment variables, including names containing OPENAI, HF_ and FOUNDRY, reads other processes' environments, takes credential files such as Azure CLI sessions, .npmrc, .git-credentials and Hugging Face tokens, and requests Azure managed-identity and Azure DevOps OIDC tokens, Wiz reports. The second release added GitHub Actions OIDC token theft and hid its trigger under new file names, so a search for the first release's files will miss it. The legitimate framework code was left in place, so applications kept working, SecurityWeek reported.

Wiz's account of the attacker#

Wiz says the attacker appears to have compromised a client certificate and private key for Microsoft's Engineering System Release Pipeline (ESRP), the internal service that signs and publishes the package. A first publish attempt from AWS was rejected by the pipeline's IP filtering; a second, from Microsoft-owned IP space, succeeded. How the credentials were obtained is not known.

The tooling identified itself as Cairn, an open-source AI penetration-testing framework, and before publishing, the attacker ran the stealer on their own machine, which captured what Wiz says appears to be an AI coding-agent configuration backed by a GLM-family model. Wiz could see this because the attacker reported progress to the same public webhook it used for exfiltration. That endpoint's free tier stored 50 requests; Wiz says four came from the attacker, 29 from sandboxes, scanners and tests, and 17 represented three unique successful exfiltrations of victim data.

Claims and checks#

Microsoft has acknowledged the incident: a maintainer confirmed in a GitHub discussion that the releases were removed and an investigation was under way, SecurityWeek reported. OX Security, Wiz, Endor Labs and SafeDep analysed the malware independently, according to SecurityWeek. The signing-pipeline explanation is Wiz's; SecurityWeek's report, written before Wiz's root-cause update, said the publishing method was unknown and noted SafeDep had found a PyPI maintainer account created on the day of the attack whose involvement was unconfirmed. The pattern of stolen AI and cloud keys as the target matches the LiteLLM compromise in October, where Wiz said stolen credentials were checked for Bedrock access within hours.

Wiz's advice is to search lockfiles, CI logs and images for 1.26.1 or 1.26.2, pin to 1.26.0, check for the dropped files, and rotate every secret reachable from an affected host or runner. Stolen managed-identity and OIDC-exchanged tokens are valid for up to 24 hours, it says.

Limits#

  • Microsoft has not published its own account; the certificate explanation and the exfiltration counts are Wiz's, and Wiz could see only the requests the webhook stored.
  • The identification of Cairn and of a GLM-backed coding agent rests on the attacker's own tooling labels, as read by Wiz; we have not seen evidence of who the attacker is.
  • We have not seen a Microsoft statement on whether other packages published through the same pipeline were affected; SecurityWeek noted other Microsoft repositories had disappeared from GitHub but said any link was unclear.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading