Update: agent-framework attacker held a Microsoft release-pipeline key and ran an AI pentest tool, Wiz says
Wiz says the poisoned PyPI releases were pushed with a key for Microsoft's internal release service and three victims lost data. Microsoft hasn't confirmed. Keep versions pinned.
Wiz Research now says the two poisoned agent-framework releases published to PyPI on November 17, 2026 appear to have been pushed with an illicitly obtained certificate and private key for ESRP, the internal service Microsoft uses to sign and publish packages, by tooling that called itself Cairn, an open-source AI penetration-testing framework. Wiz counts three confirmed victim data thefts; Microsoft has not confirmed the findings. If you installed 1.26.1 or 1.26.2, our original guidance stands, and every user of Microsoft's Python agent packages should pin versions and hold off on brand-new releases until Microsoft says the publishing path is clean.
What happened, in plain English
Microsoft Agent Framework is a toolkit developers use to build AI agents — programs that use a large language model (LLM) to plan steps and call tools. On November 17, two fake versions of its Python package appeared on PyPI, the public Python package library. They worked normally but secretly collected passwords, tokens and cloud keys and sent them to the attacker. Both were removed within about an hour. We covered the incident on November 18.
What was missing then was how the attacker got permission to publish under Microsoft's name. In an update published at 14:00 UTC on November 18, Wiz says the attacker appears to have held a certificate and private key for Microsoft's Engineering System Release Pipeline (ESRP). A first publishing attempt from Amazon Web Services was blocked by ESRP's IP address filter; a second attempt, from Microsoft-owned IP space, succeeded. How the attacker obtained the certificate is not known.
An analogy: in our first story the question was whether a thief had copied the key to one shop. Wiz's evidence suggests the thief may have had a pass to the warehouse that supplies many shops — although so far only one shop is known to have received bad goods.
Wiz could see part of the operation because the attacker's own tools reported their progress to the same public collection address used for the stolen data. That is also where the AI angle appears: the tooling called itself Cairn, and a test run on the attacker's own machine captured settings for an AI coding agent.
Are you affected? What to do now
If you installed agent-framework 1.26.1 or 1.26.2 (live roughly 20:35–21:30 UTC on November 17), nothing about the response has changed: rebuild on 1.26.0 and rotate every secret the machine or CI runner could reach. Wiz's update adds that only three successful victim exfiltrations were captured — but the attacker's free webhook.site endpoint logged at most 100 requests, so later attempts may have been rejected rather than never made. Absence from that count is not proof you are safe.
If you use other Microsoft Python agent packages, there is no evidence of malicious releases beyond the two versions above. But two reports point to a publishing path that may not yet be closed:
- Pin exact versions with hashes for
agent-frameworkand related packages, and avoid floating version ranges in CI until Microsoft confirms the root cause is fixed. - Delay new releases. SafeDep reported on November 18 that a maintainer account which joined PyPI on November 17 still had upload rights to
agent-frameworkand five sibling packages:agent-framework-openai,-purview,-redis,-toolsand-postgres. None of those has had a release since November 12. SafeDep advises holding off on any new release until Microsoft removes that account and confirms releases are clean. - Treat a release without a source distribution or GitHub tag as suspicious. Both malicious versions were wheel-only, with no matching tag or release manifest in the repository, according to the GitHub issue and Wiz.
- Hunt with Wiz's published indicators. Wiz lists the exfiltration URL on webhook.site, the user-agents
af-runtime-diagnostics/1andruntime-diagnostics/1, the startup filesagent_framework_meta.pthandaf_runtime_diag.pth, the environment variable_AF_TEL=1, and SHA-256 hashes for both malicious wheels. In proxy logs, a 200 response suggests the upload was accepted; 410 or 429 suggests it was rejected. - Check Azure and GitHub identity logs. Stolen Azure tokens can stay valid for up to 24 hours. For GitHub Actions jobs that installed 1.26.2 with
id-token: write, treat the OIDC (OpenID Connect) tokens as exposed and review Microsoft Entra ID federated credentials.
If you do not use Microsoft Agent Framework for Python, there is nothing to do.
The expert view
Root cause, revised but unconfirmed. On November 17 the working theory — from the Upwind researcher who opened the GitHub issue — was a compromised PyPI upload credential. Wiz's update moves the problem upstream into Microsoft's own release infrastructure. The detail about the IP allow-list matters: the attacker had a valid credential, was stopped by a network control, then came back from address space that the control trusted. That is a classic failure of controls that check where a request comes from instead of what is making it. It also hints that the attacker either had a foothold inside Microsoft-owned cloud ranges or could rent compute there — Wiz does not say which.
The two research accounts may describe the same path from different ends. ESRP is the only production publisher for this repository, according to the project's own documentation cited in the issue, while SafeDep sees a new PyPI maintainer account attached to the project. Neither Microsoft nor PyPI has said which credential actually performed the uploads, or whether they are connected.
Why it matters beyond one package. A release-pipeline credential is more valuable than a single project's token: Wiz describes ESRP as the service Microsoft uses to sign and publish packages, not a tool dedicated to this one project. Wiz does not claim any other package was touched, and there is no evidence of that. But until Microsoft states the certificate's scope and confirms it has been revoked, defenders cannot rule it out. That is the strongest reason to keep strict pinning in place.
The AI angle — read it carefully. The attacker's tooling labelled itself Cairn and staged the package in a working folder of the same name. Cairn is an open-source framework that lets AI agents carry out penetration tests; Unit 42 reported a suspected China-nexus group using it earlier this month. The attacker's own test run also captured an AI coding agent configuration that Wiz believes is the open-source Pi harness, driving a GLM-family model through a local proxy. Because Cairn is freely available, the name says nothing about who is behind this incident. Wiz makes no attribution, and neither do we. What it does show is that agent tooling is now part of ordinary supply-chain operations — and that it can be noisy: it reported its progress to the same endpoint that defenders were watching.
Severity in context. There is no CVE or CVSS score, because this is a malicious release, not a code flaw. The realised damage looks limited: under an hour of exposure, three confirmed victims, and a capped free-tier collection service. The potential damage depends on one unanswered question: what else the stolen signing credential could publish.
Still unknown: how the ESRP certificate was obtained and whether it is revoked; whether the SafeDep-flagged account was removed; whether any credentials taken from the three victims were used; and whether Microsoft will issue a formal advisory. As of November 19, the only Microsoft statement remains a maintainer's November 17 note that the versions were pulled and an investigation is under way.
Official sources
- Wiz Research — Microsoft agent-framework compromised, updated November 18, 2026 with root cause and exfiltration analysis
- microsoft/agent-framework GitHub issue #10411 — original report and Microsoft maintainer confirmation (November 17, 2026)
- SafeDep — Microsoft agent-framework compromised twice on PyPI (updated November 18, 2026)
- Coverage: SecurityWeek
- Coverage: The Hacker News
- Earlier on AI Cyber Brief: Microsoft's agent-framework package on PyPI was hijacked to steal Azure and CI credentials