The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted codebases and external data sources, creating a potential surface for indirect prompt injection if the audited files contain malicious instructions.
Ingestion points: Target repository source code, build manifests, test outputs, and third-party dependency data analyzed across all hunting phases (documented in RECONNAISSANCE.md and SKILL.md).
Boundary markers: The skill uses explicit instructions to sub-agents to treat all target content as untrusted and implements strict "Write isolation" and "Universal execution safety" rules in SKILL.md.
Capability inventory: The system can execute shell commands, build tools, and local test harnesses within an OS-enforced sandbox (referenced in HUNTING.md).
Sanitization: Audit findings are stored in structured JSON and validated by zero-dependency Node.js scripts (validate-findings.cjs) to prevent data-driven attacks on the parent agent.
[COMMAND_EXECUTION]: The skill provides a structured workflow for executing code from the target repository to verify potential vulnerabilities.
Evidence: SKILL.md and HUNTING.md detail the use of sandboxed local execution for "Candidate validation." The instructions enforce strict containment, including no external networking, read-only target mounts, and resource quotas to mitigate the risks of executing potentially malicious audited code.
[EXTERNAL_DOWNLOADS]: The skill's documentation provides installation instructions via a public repository.
Evidence: README.md references the official repository at github.com/cloudflare/security-audit-skill. As Cloudflare is a trusted organization, this reference is informative and does not escalate the security risk.

cloudflare-security-audit