The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is specifically designed to audit codebases (untrusted data) and incorporates robust defense-in-depth measures. It mandates the use of isolated sub-agents (hunters and verifiers) and requires all target-controlled code execution to occur within an OS-enforced sandbox with no network access and strict resource limits.
[DATA_EXFILTRATION]: The instructions explicitly forbid contacting external endpoints, registries, or shared services during the audit. It uses a 'write isolation' strategy where agents can only write to a temporary scratch directory, and a secure 'Artifact promotion procedure' is used to move verified, non-sensitive results to retained storage.
[REMOTE_CODE_EXECUTION]: The skill prevents remote code execution by requiring that all tools and dependencies used during the audit be available locally, and by disabling network access for the auditing sandbox.
[DYNAMIC_EXECUTION]: The skill includes two JavaScript utility scripts (validate-coverage-ledger.cjs and validate-findings.cjs) used to verify the structure and safety of audit artifacts. These scripts are local, transparent, and implement security best practices such as using race-safe file descriptor operations and enforcing strict input limits.

security-audit