The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external codebase data which could contain malicious instructions designed to influence the agent's findings or behavior during the audit.
Ingestion points: The skill ingests untrusted content from target codebases via the Read, Glob, and Grep tools during the Reconnaissance and Hunting phases defined in SKILL.md.
Boundary markers: The skill documentation (SKILL.md) mandates process-level boundaries, including the use of isolated sub-agents (agent-teams) and a 'finder ≠ validator' requirement where a separate agent must attempt to disprove findings.
Capability inventory: The skill environment includes Bash and Task for shell command execution, along with file system access tools (Read, Glob, Grep), as specified in the SKILL.md frontmatter.
Sanitization: The skill contains an explicit anti-pattern warning in SKILL.md: 'Read and analyze; never execute untrusted target code to "see what it does"', which serves as a critical directive to the agent to avoid dynamic execution of untrusted payloads.

security-audit