youtube-ingest

Warn

Audited by Gen Agent Trust Hub on Jul 30, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses shell commands to interact with yt-dlp and git. Passing the user-provided <youtube_url> directly into a shell command string (uvx yt-dlp --get-title "<youtube_url>") creates a risk of command injection if the input is not sanitized by the platform.\n- [DATA_EXFILTRATION]: The skill contains a hardcoded absolute path to a specific user's directory: /Users/mtm/Documents/Obsidian Vault/. Using absolute paths with specific usernames exposes local system details and limits the skill's portability and security.\n- [PROMPT_INJECTION]: The workflow relies on external, untrusted metadata from YouTube to generate filenames, commit messages, and note content, which presents a surface for indirect prompt injection.\n
  • Ingestion points: Video titles and summaries fetched via yt-dlp and WebFetch in SKILL.md.\n
  • Boundary markers: The skill lacks explicit markers or instructions to the agent to treat fetched metadata as untrusted data.\n
  • Capability inventory: Includes shell command execution (yt-dlp, git), local filesystem writes, and interaction with the social-to-imgur skill.\n
  • Sanitization: While Step 3 provides some character stripping for filenames, the summary and title are used in git commits and note bodies without escaping or validation.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 30, 2026, 06:08 PM
Security Audit — agent-trust-hub — youtube-ingest