Auto-Empirical Research Skills vendors many third-party skill repositories. Treat every new or updated skill as executable instructions for an AI agent, not as ordinary prose.

Security reports are accepted for:

• Malicious or surprising instructions in SKILL.md, references, hooks, scripts, or bundled assets.
• Credential exfiltration, reverse shells, arbitrary network callbacks, unsafe shell execution, or hidden payloads.
• Prompt-injection patterns that ask an agent to ignore user, system, or repository safety constraints.
• Misleading install instructions that fetch unreviewed code.

Please open a private security advisory if GitHub enables advisories for this repository. If not, open a minimal public issue that states "security report available" without exploit details, and the maintainers will coordinate a private channel.

Include:

• Affected path and commit SHA.
• Exact suspicious lines or files.
• Why the behavior is unsafe.
• Whether the issue comes from upstream or from AERS packaging.

We aim to triage credible reports within 7 days. High-risk findings may result in temporarily removing or disabling a skill entry until the upstream source is fixed or clearly documented.

• Prefer skills with clear licenses, minimal dependencies, and local/offline execution paths.
• Review hooks and scripts before enabling them in an agent runtime.
• Do not run vendored code with secrets in the environment unless the skill explicitly needs them and you trust the source.
• Re-run make validate after local edits.

• Dependabot monitors GitHub Actions updates through .github/dependabot.yml.
• OpenSSF Scorecard runs weekly and on pushes to main, with SARIF uploaded to GitHub code scanning.
• make validate includes a workflow policy check for explicit permissions, non-persistent checkout credentials, no pull_request_target, and no downloaded-script pipe-to-shell patterns.

• SECURITY-SCAN-REPORT.md — deep audit of the curated in-repo skills (13 grep threat classes, full manual hook review, parallel agent content review).
• docs/archive/SECURITY-SCAN-REPORT-23k-archive.md — historical archive (2026-07-06). Breadth scan of the 23,000+-skill upstream pool (95 repositories, 62,957 files, 16 automated threat patterns plus sampled triage and full hook re-review). Superseded by the current SECURITY-SCAN-REPORT.md after the catalog was trimmed to 1,150 skills.