Edition of Thursday 19 November 2026

Microsoft agent-framework Compromised Twice on PyPI

Comments on “Microsoft agent-framework Compromised Twice on PyPI”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

SafeDep reports that agent-framework 1.26.1 and 1.26.2, Microsoft's Python metapackage on PyPI, carried a credential stealer. A .pth file runs it at every Python start, so installing is enough. It takes environment variables and credential files, requests fresh Azure DevOps, managed identity and, in 1.26.2, GitHub Actions tokens, and posts them to webhook.site. PyPI removed both; 1.26.0 is the last clean release. A maintainer account created that day may be the uploader, which is unconfirmed. Treat reachable secrets as exposed and hash-lock the dependency tree.

Read the original at safedep.io. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 2 of 144 with a score of 0.82. The panel rated it highest on practical relevance (0.94) and technical depth (0.93), lowest on new information (0.59). The judges disagreed most on new information (spread 0.87), so the story is marked contested. Placed No. 2, No. 1 by score alone: close to “Microsoft's agent-framework on PyPI shipped a credential stealer in two releases”, No. 1 on Wednesday 18 November 2026 (−0.049). The placement terms move stories; they are not part of the score.

AI judgeJudge 2claude-haiku-5-5
Detailed, well-substantiated supply-chain malware analysis of a major Microsoft PyPI package with IoCs and remediation; highly actionable for builders, with fresh info not in prior week's list.
AI judgeJudge 3gpt-5.6-luna
A highly actionable supply-chain compromise report with code-level evidence, indicators, affected dependency paths, credential theft mechanics, and remediation guidance. It substantially overlaps a listed recent story, reducing new-information probability.

Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.

Score

0.82total under rubric v2.3Contestedjudges differ by 0.87 on new information

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.76
  • Reader interest0.92
  • Audience breadth0.70
  • Practical relevance0.94
  • Substantiated0.92
  • Technical depth0.93
  • New information0.59
  • Counts against: Promotional0.03
  • Counts against: Overclaim0.13

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.740.720.820.760.10+0.24+0.182
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.900.930.930.920.03+0.26+0.239
Audience breadthHow many people who build technology does it concern?0.600.780.720.700.18+0.13+0.091
Practical relevanceHow useful is it to someone who builds software or tech products?0.980.950.900.940.08+0.22+0.208
SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations.0.920.930.900.920.03+0.05+0.046
Technical depthHow technically substantive is it?1.000.960.820.930.18+0.05+0.046
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.730.950.080.590.87 , above the contested threshold+0.05+0.029
PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional.0.040.040.010.030.03−0.19−0.006
OverclaimDo the headline or conclusions claim more than the body supports?0.170.100.120.130.07−0.11−0.014
Sum of contributions+0.822
Claim-check penalty0.000
Total, as published0.822

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.

Ranked #2 on Rubric, 19 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-19/2/"><img src="https://rubric.hrih.si/badge/2026-11-19/2.svg" alt="Ranked #2 on Rubric, 19 Nov 2026" width="238" height="28"></a>

Claim check

Claim check: 2 matches source · 3 no source given · 3 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. No source givenDate

    SafeDep flagged agent-framework version 1.26.1 22 minutes after it reached PyPI on 17 November 2026, and flagged 1.26.2 four minutes after upload.

    The article gives no link or citation for it.

    The article says

    “SafeDep flagged version 1.26.1 22 minutes after it reached PyPI on 17 November 2026”

    The source says

    No source to compare with.

  2. Couldn’t checkNumber

    PyPI lists two maintainers for agent-framework, microsoft and vibiret, and the project gets about 192,000 downloads a month.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “PyPI lists two maintainers for the project, microsoft and vibiret. The project gets about 192,000 downloads a month”

    The source says

    No passage could be compared.

    Source cited by the article: pypi.org/project/agent-framework

  3. Couldn’t checkNumber

    Every agent-framework release from 1.20.0 through 1.26.0 shipped a wheel and an sdist, while both malicious versions are a single wheel; the 1.26.1 wheel is 10,402 bytes, up from 6,681 in 1.26.0.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “The 1.26.1 wheel is 10,402 bytes, up from 6,681 in 1.26.0”

    The source says

    No passage could be compared.

    Source cited by the article: pypi.org/project/agent-framework

  4. Matches sourceDate

    In May, durabletask 1.4.1 to 1.4.3, another Microsoft package on PyPI, was published with a stolen API token and had no matching commits or tags.

    The cited source says the same thing.

    The article says

    “In May, durabletask 1.4.1 to 1.4.3 went up with a stolen API token and had no matching commits or tags”

    The source says

    “were published on May 19, 2026 using a compromised PyPI API token. The GitHub repository was not breached: no corresponding tags or commits exist”

    Source cited by the article: safedep.io/blog/malicious-durabletask-pypi-supply-chain-attack

  5. No source givenCausal conclusion

    There is no evidence linking the durabletask and agent-framework incidents, since their payloads and infrastructure share nothing.

    The article gives no link or citation for it.

    The article says

    “The payloads and infrastructure share nothing, so there is no evidence that links the two incidents”

    The source says

    No source to compare with.

  6. Matches sourceVersion

    LiteLLM 1.107.0 in October used the same shape: a one-line .pth that imports a module which forks a collector.

    The cited source says the same thing.

    The article says

    “appeared in LiteLLM 1.107.0 in October”

    The source says

    “cook.pth contains import cook, so the stealer runs in every Python process on the host: the application, pip, IDE language servers, and CI jobs”

    Source cited by the article: safedep.io/blog/litellm-pypi-compromise-credential-stealer-backdoor

  7. No source givenCausal conclusion

    The vibiret account joined PyPI on 17 November 2026, the day both malicious versions went up, which points to it as the uploader of 1.26.1 and 1.26.2, though this is unconfirmed.

    The article gives no link or citation for it.

    The article says

    “The join date points to vibiret as the account that uploaded 1.26.1 and 1.26.2”

    The source says

    No source to compare with.

  8. Couldn’t checkCausal conclusion

    PyPI has no attestation for 1.26.0 either, so consumers could not cryptographically distinguish Microsoft's uploads from the attacker's.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “PyPI has no attestation for 1.26.0 either, so a consumer had no cryptographic way to tell Microsoft’s uploads from the attacker’s”

    The source says

    No passage could be compared.

    Source cited by the article: pypi.org/project/agent-framework

Also covered by

Elsewhere

Elsewhere: HN 3 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 19 Nov 2026, 05:47 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “Microsoft agent-framework Compromised Twice on PyPI”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.