The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted codebases, which could contain malicious instructions targeting the AI agent. This is a structural risk for auditing tools.
Ingestion points: Target repository source code and configuration files are read during the reconnaissance and hunting phases as defined in RECONNAISSANCE.md and HUNTING.md.
Boundary markers: The skill employs clear prompt separation, role-based delegation between a Parent agent and sub-agents, and strict sandboxing requirements for any executed code to prevent untrusted data from escaping the audit context.
Capability inventory: The skill can execute local commands (sandboxed), write structured reports to a specified output directory, and coordinate multiple delegated agents.
Sanitization: All agent results must be returned as structured JSON, which is subsequently validated against hardcoded schemas (report-schema.json) by custom Node.js scripts (validate-findings.cjs, validate-coverage-ledger.cjs) before being aggregated into final report artifacts.

security-audit