The Agent Skills Directory

[SAFE]: The functionality described in the skill is legitimate and fully aligned with its stated purpose as a security research and vulnerability auditing tool. It references authoritative sources such as CWE, OWASP, and CVSS to guide its methodology.
[PROMPT_INJECTION]: The skill inherently processes untrusted data from target repositories, creating an attack surface for indirect prompt injection.
Ingestion points: Codebase contents, PR details, and git diffs are processed as audit targets in SKILL.md during Phase 0 and Phase 1.
Boundary markers: Untrusted content is included in agent prompts without specific boundary delimiters.
Capability inventory: The agents utilize rg, git, LSP, and the ability to execute localized reproduction code via specialized PoC sub-agents.
Sanitization: No automated sanitization of the input code is implemented before it is analyzed by the sub-agents.
Mitigation: The risk is addressed by strict behavioral instructions, including mandates for "local-only execution," the use of "toy inputs," and the avoidance of destructive actions against external systems or services.
[COMMAND_EXECUTION]: While the skill involves generating and executing Proof-of-Concept code to verify vulnerabilities, this is restricted by instructions to maintain safety, use local fixtures, and prefer static proof over unsafe execution.

security-research