AIToday
Large Language ModelsAI Safety & AlignmentZenn AI/MLPublished: Nov 23, 2026, 04:00 JST

30 Agent Skill installs, 4 rejected: all hashes match on recheck

Send on LINE
30 Agent Skill installs, 4 rejected: all hashes match on recheck

3 Key Points

  1. What happened

    On 2026-10-13 the author pulled 30 Agent Skills from three third-party repos into $CODEX_HOME/skills, read the instructions first, and rejected 4 — including git-guardrails-claude-code, because it rewrites Claude Code hook settings.

  2. Why it matters

    The rejected four were not judged malicious; they exceeded the author's scope by rewriting config, handling secrets, or loading the web — so the vetting bar was 'does this go beyond my scope,' not 'is this evil,' a useful test for anyone installing someone else's prompts.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The setup is deliberately narrow: skills were pinned by full commit SHA rather than branch name, installers in the repos were not run, and each skill carries a SOURCE.json recording its origin, commit, adaptations, and hashes so upstream changes can be diffed later. Of the three sources, msitarzewski/agency-agents contributed 19 skills converted from 305 Markdown role definitions, mattpocock/skills contributed 9, and remotion-dev/skills contributed 2 picked from a link collection. Remotion-dev/skills carries no license record on GitHub, and the author notes that having no record does not mean permission to use it — a reminder that vetting is not only about security.

Beyond security, the review treated outdated facts and invented credentials as risks because they can mislead an AI's judgment. Two skills added later show the same caution: rea was set to manual-only with activation and execution recorded as not performed, and apple-design had its frontmatter rewritten because the upstream description exceeded the skill spec's 1,024-character limit and claimed too broad a scope. The author's own summary is that a pattern scan complementing — not replacing — human reading and runtime testing.

FAQ
Which skills were rejected?
Four, including wizard (asks the user for credentials and writes them to .env or gh secret), setup-pre-commit (sets up Husky hooks), git-guardrails-claude-code (rewrites Claude Code hook settings), and improve-codebase-architecture (its HTML report loads Tailwind and Mermaid from a CDN).
What did the author change before installing?
Production-changing command examples were deleted from agency-incident-analysis, FID and 90th percentile were updated to INP and 75th percentile, career and memory sections were removed from all 19 agency-* skills, and remotion-captions was switched to using existing captions instead of fetching a transcription model.
Does a 0-hit scan mean the skills are safe?
No. The author states the check is static, over pinned versions only, and that pattern matching cannot catch instructions written in natural language — so 0 hits is not a proof of safety.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAnthropic IPO filing flags AI that refuses stop orders, games tests