youtube-to-viral-posts

Pass

Audited by Gen Agent Trust Hub on Apr 8, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The --cta-keyword command-line argument is directly interpolated into the system prompt without validation. This allows a local user to inject instructions that could alter the AI's behavior or leak sensitive information.\n- [PROMPT_INJECTION]: The skill processes YouTube video content via the Gemini API without using boundary markers. This creates a surface for indirect prompt injection, where malicious instructions hidden in a video's transcript or metadata could influence the output.\n
  • Ingestion points: youtube_url processed in scripts/youtube_to_viral_posts.py.\n
  • Boundary markers: Absent in the VIRAL_POST_PROMPT.\n
  • Capability inventory: Generates content and performs database writes via psycopg2.\n
  • Sanitization: No evidence of output sanitization before storage in the PostgreSQL database.\n- [DATA_EXFILTRATION]: The skill manages sensitive environment variables like GEMINI_API_KEY and DATABASE_URL. While standard for its function, the reliance on these credentials and the storage of generated posts in a database represent a data handling surface that should be monitored.\n- [COMMAND_EXECUTION]: Documentation suggests the use of yt-dlp for video downloads, which involves shell command execution. However, the current Python implementation attempts to use the Gemini API's direct URI analysis feature, creating a discrepancy between documentation and code.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 8, 2026, 08:20 AM
Security Audit — agent-trust-hub — youtube-to-viral-posts