The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes an architecture that ingests untrusted data from multiple sources (files, web search, logs, screenshots) and processes them as 'percepts'. This creates an indirect prompt injection surface where attackers could embed instructions in external data.
Ingestion points: Uses read_file and search_web to collect data, and references logs and screenshots as perception inputs (SKILL.md).
Boundary markers: The skill explicitly recommends implementing a "rigorous sanitization boundary" and using "cryptographically verified environment state" (SKILL.md).
Capability inventory: The skill lists write_to_file, run_command, and replace_file_content as available capabilities for the agent (SKILL.md).
Sanitization: The skill does not provide implementation code for sanitization but provides high-level guidance and warnings for its implementation in the "Security & Guardrails" section.

sensory-motor