Would a Maintainer-Based Risk Score Have Caught the agent-framework PyPI Compromise?
No. We sell one, so here is why not, and what would have worked.
No. On November 17, 2026, two malicious releases of Microsoft’s agent-framework package, 1.26.1 and 1.26.2, went up on PyPI. Nothing about the package’s maintainers looked wrong before or after. Scores built on maintainer count, publisher activity and release cadence, ours included, rate who can ship the next version. They cannot tell a stolen upload from a real one when the package itself is healthy.
What the maintainer signals said
Read from PyPI on November 19: two accounts are listed as maintainers, one of them named microsoft. The package had shipped 1.25.0 on November 7 and 1.26.0 on November 12, and had 39,523 downloads in the past week (pypistats). That is an active, multi-publisher project. It is not the single-maintainer, gone-quiet profile that maintainer scores are built to flag.
Provenance did not separate them either. PyPI shows no attestations on 1.25.0 or 1.26.0, so a release without one looked the same as the legitimate releases before it.
What did catch it
Tools that look at the release itself. Elastic reports 1.26.1 went live at 20:35 UTC and their monitor flagged it at 20:41. The new release added a .pth file, which Python runs at every startup without the package ever being imported. 1.26.2 followed at 21:08 UTC with the same stealer under new file names, and Elastic says PyPI removed both within the hour. StepSecurity, Wiz and Endor Labs have the indicators and the clean-up steps. Use their pages if you installed either version.
What would have kept it out
A release-age delay. Endor Labs puts it plainly: even 24 hours of minimum release age would have kept both versions out. pip and uv both support this now:
pip install --uploaded-prior-to P3D agent-framework
# pyproject.toml
[tool.uv]
exclude-newer = "3 days" We ran both on November 19 with pip 26.3 and uv 0.13.9. Each resolved agent-framework to 1.26.0. A pinned lockfile with hashes does the same job for builds that never re-resolve.
Where Commit fits
Commit now shows the package as compromised, because it reads the OpenSSF malicious-packages feed through OSV:
npx proof-of-commitment --pypi agent-framework On November 19 that printed “COMPROMISED — MAL-2026-20101: malicious versions 1.26.2, 1.26.1”. OSV published that advisory at 22:12 UTC, about 97 minutes after 1.26.1 went live. So Commit helps you answer “did we install a bad version?” after the fact. It would not have stopped the install. Maintainer scores are for a different failure: a package whose only publisher has gone quiet, where nobody would notice a strange release. Use both kinds of check, and a release-age delay on top.
Commit is our tool, so weigh what this page says about it accordingly. Every number here was read on November 19, 2026; re-run the commands rather than trusting this page.