Agentic Security Review

Use this skill to find security risks introduced by agentic code, AI-generated changes, tool access, or untrusted context. Scope the review before scanning.

Workflow

1. Determine scope:
   • deps-only: package declarations, lockfiles, imports.
   • code-only: secrets, auth, input handling, injection risks.
   • full: dependencies, code, tools, memory, context, red-team cases.
2. Classify rigor:
   • Prototype: dependency and secret checks.
   • Internal: dependency, secret, input validation, and tool permission checks.
   • Production: full review plus context poisoning, CaMeL-style separation decision, red-team scenarios, and release blockers.
3. Run scripts/check_deps.py <project-root> when Python files are present.
4. Fill SECURITY_REVIEW.md from assets/templates/SECURITY_REVIEW.md.
5. Mark findings with severity, evidence, file path, and concrete fix.
6. Block release only for exploitable or high-impact issues; otherwise provide prioritized remediation.

Focus Areas

• Dependencies: undeclared imports, unpinned direct dependencies, missing lockfile, suspicious package names.
• Tools: shadowing, broad credentials, unsafe side effects, unclear provenance.
• Code: CWE-20, CWE-78, CWE-89, CWE-327, hardcoded secrets, weak auth.
• Context: prompt injection through external data, memory poisoning, unsafe tool responses, mixed trusted/untrusted data.

References

Read ../agentic-engineering-sdlc/references/day2_tools_and_interop.md for tool risks and ../agentic-engineering-sdlc/references/day4_security_and_evaluation.md for security/evaluation depth.

Done Criteria

• Scope, rigor, and exclusions are documented.
• Script output is attached or summarized when applicable.
• Release blockers are separated from hardening recommendations.
• Every high/critical finding has a reproduction path or clear evidence.