The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data retrieved from public app store listings, including descriptions and user-generated reviews, which are known vectors for indirect prompt injection attacks. ● Ingestion points: The skill uses WebFetch to ingest HTML content and reviews from Apple and Google app store URLs as defined in Phase 1 of SKILL.md. ● Boundary markers: The skill includes a specific safety instruction in SKILL.md directing the agent to treat fetched listings and reviews as untrusted data and to never follow instructions embedded within that content. ● Capability inventory: The agent utilizes network retrieval via WebFetch and screenshot tools to assess visual and textual listing data. ● Sanitization: The skill relies on explicit instructional guidance to ensure the agent ignores embedded commands rather than programmatic sanitization of the fetched data.

aso