fountain-clip-finder

Pass

Audited by Gen Agent Trust Hub on Sep 28, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The Python script modules/external-source/scripts/read-segments.py executes external command-line tools to process video and audio data.\n
  • It uses subprocess.run to call yt-dlp, ffmpeg, and ffprobe for metadata extraction and transcription tasks.\n
  • User-provided URLs and file paths are passed as arguments to these tools. While shell=True is not used, this remains a significant capability.\n- [EXTERNAL_DOWNLOADS]: The skill uses yt-dlp to fetch caption tracks and metadata from YouTube watch pages.\n
  • Evidence: download_captions and fetch_caption_words functions in modules/external-source/scripts/read-segments.py. YouTube is a well-known service, making this a legitimate operation for the skill's purpose.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript content from external sources (YouTube or local files) which is then used to generate social media copy, posing a risk of indirect prompt injection.\n
  • Ingestion points: Transcripts are read from YouTube or local files in modules/external-source/scripts/read-segments.py and passed through the discovery and boundaries modules.\n
  • Boundary markers: The modules/copy/MODULE.md instructions do not specify the use of clear delimiters or instructions to ignore embedded commands when interpolating transcript text into prompts for generating labels and post text.\n
  • Capability inventory: The skill can execute shell commands via subprocess (in the external-source module) and publish draft posts via the Fountain Social API.\n
  • Sanitization: A "safety pass" is mentioned in modules/copy/MODULE.md for content risks, but no specific sanitization for prompt injection is described for the ingested transcript data.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 28, 2026, 06:12 AM