The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted data from external websites via WebFetch and uses that data to influence its internal logic and source code recommendations.
Ingestion points: WebFetch operations in Fase 1 targeting external URLs (Meta Ad Library, Pinterest, blogs, etc.).
Boundary markers: Absent. The instructions do not include delimiters or warnings to ignore instructions embedded within the fetched content.
Capability inventory: The agent has the ability to read local source code files (src/lib/ai/prompts/*.ts), write to the filesystem (recursos/ad-intelligence/), and perform network operations (WebFetch).
Sanitization: Absent. The skill instructs the agent to extract "ALL" insights and be "extremely detailed" without any validation or filtering of the content.
[EXTERNAL_DOWNLOADS]: The skill performs automated network requests to various external domains (foreplay.co, adcreative.ai, propphy.com, etc.) to fetch marketing data. While these are presented as research targets, fetching content from arbitrary URLs based on search results is a standard risk vector for data ingestion.

actualizar-imagen