The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection through its 'Live Research Pass' feature.
Ingestion points: The skill instructions in SKILL.md direct the agent to fetch and process data from external websites, videos, and other research-focused skills (e.g., deep-research, watch-video).
Boundary markers: The skill does not prescribe the use of explicit boundary markers or delimiters when interpolating data from these external sources into the active prompt context.
Capability inventory: The skill is primarily focused on text generation and persona simulation. It includes instructions to write custom dossiers to a local directory (.agents/advisors/), which is a legitimate persistence feature for custom advisor configurations.
Sanitization: The skill relies on instructional 'Grounding Rules' (e.g., 'No fabricated quotes') to maintain output fidelity, but it lacks programmatic sanitization of external input. Furthermore, SKILL.md explicitly instructs the agent to 'trust the research' if external findings contradict internal dossiers, which could lead the agent to adopt malicious instructions embedded in poisoned web content.

marketing-council