The Agent Skills Directory

[PROMPT_INJECTION]: The instructions in SKILL.md include defensive measures to explicitly ignore and flag prompt injection attempts (such as 'ignore previous instructions') that may be encountered within the untrusted source code being analyzed.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as its primary function is to process untrusted source code.
Ingestion points: Analyzed source code files, context slices, and repository metadata as defined in SKILL.md.
Boundary markers: An 'Untrusted-content envelope' is defined in SKILL.md to isolate target data from instructions.
Capability inventory: The skill uses various tools including raptor-audit for recording, raptor-study-loop for external model interaction, and raptor-run-sandboxed for execution, as detailed in SKILL.md and review.md.
Sanitization: All hypotheses generated by the LLM must be validated by mechanical tools (Semgrep, SMT, Coccinelle) before being emitted as findings.
[DYNAMIC_EXECUTION]: The skill allows for the creation and compilation of minimal reproducers using gcc and other tools via the raptor-run-sandboxed utility, which is a legitimate requirement for verifying security vulnerabilities.

audit