The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from YouTube videos, which could contain malicious instructions designed to influence the agent's behavior during content generation.
Ingestion points: Untrusted content is fetched from YouTube via youtube_video_transcripts_fetch, youtube_videos_read, and youtube_videos_search_top in SKILL.md and several scripts including analyze_thumbnail_concepts.py, generate_seo_description.py, and generate_seo_titles.py.
Boundary markers: There are no explicit delimiters or specific instructions for the AI to ignore embedded commands within the fetched transcript or summary text when it is interpolated into prompts for ai_functions_run or images_generate.
Capability inventory: The skill possesses capabilities to generate new text assets (titles, descriptions, social posts) and images (thumbnails) based on the ingested content.
Sanitization: The implementation lacks mechanisms to validate or filter the external YouTube content before it is passed to the AI models for processing and generation.

youtube