The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs local environment setup and coordinates with other skills through command-line invocations.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes content from untrusted documentation and skill files through sub-agents. * Ingestion points: The skill reads local files from user-defined paths like docs/ or skills/ in Step 4 of SKILL.md. * Boundary markers: The prompt template in Step 2 uses structural headers such as CONTEXT, MANDATORY CONSTRAINTS, and YOUR TASK to define the data boundaries. * Capability inventory: The skill can spawn general-purpose sub-agents, write findings to the .agents/ directory, and execute the council skill. * Sanitization: There is no evidence of sanitization or escaping of the target file content before it is interpolated into the prompt templates.

red-team