The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute various local scripts in the libexec/ directory (such as raptor-validation-helper, raptor-validate-schema, and raptor-run-feasibility) and to compile proof-of-concept exploits using gcc. These actions are necessary for its primary purpose of vulnerability validation and feasibility analysis.
[PROMPT_INJECTION]: The skill possesses a significant ingestion surface for indirect prompt injection because its core function is to process and analyze untrusted external data, including target repository source code and previous vulnerability findings. It mitigates this risk through a multi-stage validation process and strict instructional gates.
Ingestion points: Target repository source code files (at TARGET_PATH), checklist.json, and findings.json.
Boundary markers: The instructions mandate strict 'Must-Gates' (such as GATE-4 NO-HEDGING and GATE-8 POC-EVIDENCE) and include a dedicated 'Sanity Check' stage (Stage C) to verify all findings against the ground-truth source code character-for-character.
Capability inventory: Execution of Python snippets, compilation of binaries via gcc, and execution of pre-approved local shell scripts.
Sanitization: Employs Stage C for manual code verification and uses automated schema validation tools to ensure output integrity.
[EXTERNAL_DOWNLOADS]: The skill's documentation contains a reference link to the author's repository (github.com/gadievron/exploitation-validator) for methodological details. This reference to the vendor's own resource is informational and does not involve the automated download or execution of remote code.

exploitability-validation