The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it ingests data from project context files (e.g., '.agents/product-marketing.md') and user-defined advisor files ('.agents/advisors/'). Ingestion points: Reads local project marketing context and user-defined advisor dossiers. Boundary markers: Explicitly requires a simulation disclaimer at the start of output and enforces strict grounding rules to prevent fabricated quotes. Capability inventory: Primarily generates text analysis; can call external research tools if installed. Sanitization: Instructs the agent to paraphrase positions rather than using unverified quotes, reducing the likelihood of executing embedded instructions.
[INDIRECT_PROMPT_INJECTION]: The 'Live Research Pass' feature relies on fetching external data via web search or research skills. This creates a dependency on the safety of third-party content which could contain adversarial prompts.

marketing-council