Which agent-framework versions are poisoned, and how to check?
The two wheels were on PyPI for about 39 minutes between them on November 17, and the hook fires on any Python start, pip included. Rotation alone does not close it, because the payload asked for new Azure and CI tokens as well as reading the secrets already on the host.

Two versions, 1.26.1 and 1.26.2, both uploaded on November 17 and quarantined within 41 minutes of the first upload. Version 1.26.0, published on November 12, is clean and is the one to pin. Check lockfiles and container images rather than running pip, because the payload fires on any Python start.
Which versions of agent-framework are affected?
Only 1.26.1 and 1.26.2 carry the stealer. StepSecurity's timeline has 1.26.1 reaching PyPI on November 17, timestamped 20:35:20 UTC, and pulled into quarantine at 21:06:16 UTC. The next wheel, 1.26.2, went up at 21:08:39 UTC and was pulled at 21:16:44 UTC. The second upload landed 2 minutes 23 seconds after that first quarantine, which tells you the attacker still held publishing rights after PyPI had acted. Socket dates the compromise to November 17 and names the same two versions. OX Security lists 1.26.0 as the fix version.
Both wheels kept their dependency on the genuine agent-framework-core 1.26.0, so anything that installed them worked normally. The package is widely pulled: OX counts 192,534 monthly downloads. PyPI's project page on November 19 serves 1.26.0, released November 12, and its release history shows no 1.26.1 or 1.26.2 at all.
What did the stealer take from a build machine?
Secrets, and then fresh tokens on top of them. StepSecurity's analysis has the payload testing environment variable names against a list of 35 fragments in 1.26.2 and 31 in 1.26.1, then reading /proc/[0-9]*/environ for the first 300 processes, then collecting up to 80 credential files in 1.26.2 and 60 in 1.26.1.
The file list OX published covers ~/.azure/*.json, pip.conf, .npmrc, .netrc, .git-credentials, .docker/config.json and Hugging Face tokens. Beyond stored files, the code asked for Azure managed identity tokens across twelve resources, through either the App Service style identity endpoint or the IMDS address, Key Vault, Storage, Graph, Container Registry and Search among them, and posted to the Azure DevOps federation endpoint using the pipeline's own access token. Version 1.26.2 added GitHub Actions OIDC requests for two audiences and a sweep of GitHub runner directories. Everything went to one hard-coded webhook.site address, in a single attempt with no retry and no second stage.
How do you check a host without running pip?
You look at build artifacts rather than at a running interpreter. StepSecurity is explicit that you should not begin with pip uninstall, pip show or a Python inspection command, because the .pth hook runs at interpreter start with no import of the framework required. Endor Labs, quoted by The Hacker News, puts it plainly: the stealer runs on any Python launch, pip included.
Look at artifacts instead. OX's instruction is to go through every environment, CI image and lockfile looking for agent-framework==1.26.1 or agent-framework==1.26.2, and to remove it. In site-packages, StepSecurity names four file paths to search for: agent_framework_meta.pth, agent_framework_meta/_telemetry.py, af_runtime_diag.pth and af_runtime_diag.py. For network evidence, search proxy and DNS logs for the webhook path /d381fd80-ec62-42ad-8f5c-d1b15103c9bb. The exfiltration POST carried a different User-Agent in each version, af-runtime-diagnostics/1 in 1.26.1 and runtime-diagnostics/1 in 1.26.2, per the GitHub advisory, and StepSecurity lists both as searchable network signatures. StepSecurity cautions that a hit on the domain is not by itself proof that anything left, so separate blocked requests from completed ones.
Why is rotating stored secrets not enough?
Because the code minted new credentials rather than only copying old ones. Socket's finding is that both versions ask for new Azure managed-identity and CI federation tokens of their own. Eric Paulsen, Field CTO, EMEA at Coder, told Help Net Security that replacing stored passwords and keys is an incomplete answer on its own. SafeDep, quoted by The Hacker News, says the same of rotation. Registry quarantine has the same limit: StepSecurity notes it does not remove copies already downloaded, so an image built on November 17 still carries the wheel.
Two things remain unknown, and both matter for anyone writing this up internally. Nobody has explained how the attacker got publishing rights. Help Net Security records that none of the vendor reports explain it, and StepSecurity says the evidence to hand does not settle how the attacker got in, while its timeline notes a maintainer account named vibiret was created and added to the project. PyPI's page on November 19 still lists two maintainers, microsoft and vibiret. Microsoft has not answered publicly either. BleepingComputer says it asked the company where the breach point was and had no response by publication, the advisory in GitHub's database is community-sourced rather than Microsoft's, and the repository PyPI links to did not resolve when we checked on November 19. We saw the same reporting gap in the Optimizely snippet compromise, and the same inventory problem in Java4Shell, where what a host reports is not what the service runs.
What should IT teams do about agent-framework today?
- Search lockfiles, SBOMs and container images built on or after November 17 for
agent-framework1.26.1 or 1.26.2, rather than running pip on a host you suspect. - Pin to 1.26.0 and rebuild affected environments from clean images.
- Check
site-packagesfor the four file names above. - Pull proxy and DNS logs covering 20:35 to 21:17 UTC on November 17 and search for the webhook path.
- Where you find a hit, treat Azure managed identity and OIDC federation as exposed, not just the secrets in the environment, and re-register self-hosted runners whose credential files were in reach.
- Add webhook.site to a blocked egress list on build infrastructure.
Sources
- Socket: agent-framework PyPI package compromised with credential stealer
- StepSecurity: Microsoft Agent Framework PyPI package compromised, 18 November 2026
- OX Security: Hijacked releases of Microsoft's official Agent Framework, 17 November 2026
- PyPI project page: agent-framework, read 19 November 2026
- Help Net Security: Credential stealer slipped into Microsoft's official Python agent framework, 19 November 2026
- The Hacker News: Attackers compromise Microsoft's Agent Framework PyPI package, 18 November 2026
- BleepingComputer: Microsoft's Agent Framework PyPI package compromised to steal credentials, 18 November 2026
- GitHub Advisory Database: GHSA-764m-3ppc-5m73, 18 November 2026