The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted instructions from agent files in .claude/agents/. A malicious agent definition could contain text designed to influence the audit outcome or user decisions.\n
Ingestion points: Reading .md files using the Read tool during discovery.\n
Boundary markers: None specified for isolating audited text.\n
Capability inventory: Uses the Edit tool for filesystem modifications based on analysis.\n
Sanitization: Lacks sanitization of ingested content.\n- [COMMAND_EXECUTION]: The skill executes shell discovery commands with Bash using a user-provided path.\n
Evidence: Uses find commands like find .claude/agents ... with optional path arguments.

eval-agents