The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection. It retrieves transcripts, video metadata, and channel information from external, attacker-controlled URLs (YouTube, Vimeo, etc.). This content is passed to the agent for synthesis and reporting without boundary markers or explicit instructions to ignore embedded commands. Malicious actors could hide instructions in subtitles or video descriptions (e.g., using HTML comments or specific text patterns) to influence the agent's behavior during the research process.
Ingestion points: scripts/yt_research.py (via transcript and metadata subcommands).
Boundary markers: Absent; transcripts are output as plain text or SRT.
Capability inventory: Access to Bash (python3, yt-dlp, jq) and WebSearch.
Sanitization: The script performs basic character cleaning for VTT/SRT tags but does not sanitize the natural language content for potential prompt injection.
[COMMAND_EXECUTION]: The main script scripts/yt_research.py uses subprocess.run() to invoke the yt-dlp binary. While the implementation correctly uses list-based arguments to mitigate shell injection risks, the script facilitates the execution of a powerful external utility with parameters (URLs, queries, and cookie flags) derived from potentially untrusted user or network input.
[EXTERNAL_DOWNLOADS]: The skill is designed to download audio files and subtitles from external platforms to the local filesystem (defaulting to ~/youtube-research). While these downloads target well-known services, the agent should be aware that it is writing external data to the host environment.

search-youtube