The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted content from video platforms to generate research reports, creating an attack surface for indirect prompt injection.\n
Ingestion points: The agent is instructed to read and synthesize video transcripts, titles, and descriptions fetched via the transcript and metadata subcommands (defined in SKILL.md and implemented in scripts/yt_research.py).\n
Boundary markers: There are no explicit instructions or delimiters (e.g., XML tags or clear warnings) to help the agent distinguish between its own system instructions and potentially malicious instructions embedded within video transcripts.\n
Capability inventory: During the research process, the agent has access to powerful tools including Task (for spawning sub-agents), Bash (for shell execution), and WebSearch (for external queries).\n
Sanitization: While the yt_research.py script cleans technical subtitle tags (VTT/SRT), it does not sanitize the natural language content for injection patterns before passing it to the agent.\n- [COMMAND_EXECUTION]: The skill facilitates the execution of shell commands by wrapping the yt-dlp binary.\n
Evidence: The run_ytdlp function in scripts/yt_research.py uses subprocess.run() to execute commands. Although it correctly uses a list-based argument structure and avoids shell=True, it remains a mechanism for dynamic command execution based on agent-provided inputs.\n- [EXTERNAL_DOWNLOADS]: The skill is built to download data from external video hosting platforms.\n
Details: It fetches transcripts, metadata, and audio files from services like YouTube, Vimeo, and Twitter. These operations rely on the yt-dlp dependency as specified in the usage instructions and references/platforms.md.

search-youtube