The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface through the ingestion of untrusted video transcripts.
Ingestion points: Transcripts are fetched via the transcript subcommand in yt_research.py and processed by the agent during research rounds in SKILL.md.
Boundary markers: The skill does not implement delimiters or system-level instructions to ignore potential commands embedded within transcripts when they are passed to the agent context.
Capability inventory: The skill allows for command execution via yt-dlp and python3, file writing to the local filesystem (~/youtube-research), and the spawning of sub-tasks via the Task agent tool.
Sanitization: yt_research.py performs regex-based cleaning to remove VTT/SRT metadata tags, but it does not sanitize the remaining textual content for malicious prompt injection patterns or adversarial instructions.
[COMMAND_EXECUTION]: The skill facilitates the execution of external CLI tools, including yt-dlp and jq, to process video data.
Security review of yt_research.py confirms that command arguments are handled as lists in subprocess.run, which provides robust protection against shell injection attacks.
[EXTERNAL_DOWNLOADS]: The toolkit downloads metadata, audio, and subtitle files from external platforms like YouTube, Vimeo, and Twitter.
These operations are performed using the legitimate and well-known yt-dlp utility. The script includes logic to sanitize file titles to prevent directory traversal or filename-based injection.

search-youtube