Аудит безопасности Inspect AI перед запуском

security

security-audit-eval is a Claude Code skill that audits third-party Inspect AI evaluations for security risks before running them locally. It covers eight risk categories: host-side code execution (imports, `__init__.py`, postinstall hooks), sandbox configuration (Docker, `compose.yaml`, `privileged` flags), externally fetched files, supply-chain hazards (unpinned or `git+` dependencies), sandbox-breakout instructions, credential exposure, resource exhaustion, and provenance signals. The audit is strictly read-only — it does not clone repositories or execute the evaluation. Results are written to `agent_artefacts/security_audits/<eval_name>/SECURITY_AUDIT_REPORT.md` as a structured Markdown report with a clear verdict — safe, safe-with-caveats, or unsafe — backed by file:line evidence. It is intended for developers and researchers who need to vet a GitHub-hosted or local eval project before running it in their environment.