security-audit
Overview
Review concrete security boundaries and validate findings with scoped evidence, using an expanded audit workflow only when appropriate. Match guidance to the actual task and project; this document does not implement an automatic enforcer.
When to Use
Security questions, threat modeling, focused vulnerability review or an explicitly requested broader audit. A routine fix does not automatically require a multi-agent audit.
Rules & Patterns
Define the assets, trust boundaries, authorized target and relevant threat model. Trace identity, ownership, untrusted inputs and sensitive operations. Distinguish a candidate concern from a demonstrated defect.
For a finding, record the entry point, transformations, authorization decision, sink and bounded consequence. Use disposable data for reproduction. Check real filesystem containment, including link ancestors; a string prefix is not a directory boundary.
For untrusted target execution, use actual available isolation, remove credentials, limit writes and constrain network access as appropriate. A scratch directory alone is not an OS sandbox. Existing authorized development checks remain governed by the task and environment.
Use independent verification when available and authorized for a consequential audit; a second role label in one model is self-review. If tools, isolation, live services or independent review are unavailable, state the resulting evidence limit and continue useful in-scope work.
Produce artifacts proportionate to the request. A focused review can report a concise source trace and reproduction; do not create a fixed report bundle, numerical risk score or false attestation for every security question.
Code Examples
See EXAMPLES.md for an executable existing-file containment check, its symlink/prefix regressions and its concurrency/authorization limits. The example is not a substitute for a maintained race-safe filesystem boundary.
See EXAMPLES.md for scoped scenarios and evidence boundaries.
Validation Checklist
- Verify allowed and denied cases and attempt to disprove a candidate finding.
- Separate static reasoning, local reproduction, independent review and unavailable deployment checks.
Common Mistakes
Prefix-only containment; executing untrusted code with ambient credentials; presenting unverified hypotheses as confirmed vulnerabilities; claiming isolation the tools do not provide.
Integration Notes
Optional companion guidance: security, engineering-workflow. Use it only when relevant and available; the skill does not install tools or authorize external actions.
