The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions direct the agent to analyze external data such as code, configuration, and documentation. This creates a surface for indirect prompt injection if the ingested data contains instructions designed to manipulate the agent. 1. Ingestion points: SKILL.md instructions to 'Inspect the relevant code, configuration, tests, documentation, accepted proposal, or other available evidence.' 2. Boundary markers: The instructions include boundaries such as 'Read and explain; do not modify' and 'Separate observed behavior, documented behavior, proposed behavior, inference, and unknowns,' which help mitigate accidental obedience to embedded instructions. 3. Capability inventory: No additional tools or dangerous capabilities (such as shell execution or network requests) are requested or used by the skill. 4. Sanitization: No explicit sanitization, filtering, or escaping of external content is defined.

how