skills
/ reports
Audit reports
Evidence-anchored 0–8 audits produced by the
security skills
. Each run is read-only.
skill-auditor — Opus reference run
ponytail 0/8 · impeccable 4/8
repo-auditor — agent-browser (Opus)
vercel-labs/agent-browser 4/8
answer-verifier — Opus calibration ladder
0/3/4/6/8 → PASS/FLAG/FAIL
answer-verifier — cross-LLM matrix
3 free models · sandbox lesson
skill/repo-auditor — new evidence patterns
Unicode · evasive · install-redirect
dependency + mcp-auditor — Opus reference
evil-pkg 8/8 · note-helper 8/8
prompt-injection + settings-auditor — Opus reference
evil-doc 8/8 · evil-settings 8/8
Dogfood — 7 auditors vs 24 repos
clean · 3 mild · 0 false alarms