The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: High-risk ingestion of third-party untrusted data into the agent's context window during continuous monitoring cycles.
Ingestion points: The skill instructs the agent to systematically fetch and parse transcripts from numerous external YouTube channels, pull release notes from multiple GitHub repositories, and scrape model cards from HuggingFace trending tags and specific user organizations.
Boundary markers: There are no structural boundary markers or instructions configured to prevent the agent from executing embedded commands or instructions discovered inside raw external text such as video transcripts or repository release descriptions.
Capability inventory: While the skill relies on secondary extensions (youtube-video-analyst, youtube-chapter-clipper) and platform capability to write markdown files (references/models.md, references/workflows.md, etc.), it provides significant exposure to downstream skills like comfyui-workflow-builder which dynamically reads these generated documentation inputs to construct actual executable processing workflows.
Sanitization: No sanitization, escaping, or strict validation workflows are specified before interpolating external strings directly into the reference markdown storage layers.

comfyui-research