The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill audits external codebases, which presents a surface for indirect prompt injection if target files contain malicious instructions.
Ingestion points: Step 1 and 2 read repository files and git diffs into the agent context (SKILL.md).
Boundary markers: Step 3 isolates investigation by providing subagents only with relevant candidates and their dependencies (SKILL.md).
Capability inventory: The skill uses ripgrep for filesystem scanning and Task/Agent tools for parallel analysis (SKILL.md).
Sanitization: Step 4 implements a skeptic subagent pass to adversarially revalidate findings, filtering out noise or deceptive content (SKILL.md).
[COMMAND_EXECUTION]: The skill utilizes the ripgrep utility to perform read-only searches for vulnerability patterns throughout the codebase. This is a standard development tool operation.
[DYNAMIC_EXECUTION]: The workflow orchestrates parallel subagents to conduct deep-dive vulnerability investigations, representing a structured use of autonomous model capabilities for complex analysis.

security-audit