The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to execute yt-dlp commands and a local Python script. While the Python script uses list-based subprocess.run execution, which mitigates shell injection, the instructions in SKILL.md suggest passing user-provided URLs and queries directly to CLI tools without using the -- separator to terminate option parsing. This could allow for argument injection if a user-supplied input begins with a hyphen.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and displays untrusted content from external sources (YouTube video titles, descriptions, and comments). This content is presented directly to the AI agent, creating a surface for indirect prompt injection where malicious video metadata could attempt to influence the agent's future behavior.
Ingestion points: The scripts/yt_search.py script fetches and extracts fields like title, description, tags, and channel name from YouTube's API/search results.
Boundary markers: No explicit delimiters or warnings are used when presenting this external content to the agent.
Capability inventory: The agent has access to Bash and Read tools, which could be abused if the agent is manipulated by injected instructions.
Sanitization: The script performs minimal sanitization (basic newline removal in descriptions) but does not filter for instructional patterns or malicious payloads.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading audio, captions, and metadata from YouTube. It also provides instructions for installing dependencies from well-known sources, such as yt-dlp via Homebrew or GitHub, and the Deno JavaScript runtime. These references to established tools and services are documented for functional purposes.

youtube-search