universal-video-clipper
Warn
Audited by Gen Agent Trust Hub on Mar 21, 2026
Risk Level: MEDIUMCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script
ai_clip_generator.pyexecutes multiple external command-line tools using thesubprocessmodule, includingffmpegfor video editing,ffprobefor metadata extraction,yt-dlpfor downloads,whisperfor transcription, andnpx remotionfor rendering captions. While it uses list-based arguments to prevent shell injection, the reliance on multiple external binaries with broad capabilities increases the risk profile. - [EXTERNAL_DOWNLOADS]: The skill facilitates the download of video files from YouTube and potentially other platforms using the
yt-dlputility. This involves interacting with untrusted remote servers and processing potentially malicious media files. - [CREDENTIALS_UNSAFE]: In the
download_videofunction, the script attempts several download methods, two of which include the--cookies-from-browserflag for Chrome and Firefox. This command allows the tool to read the user's browser cookie database to authenticate with YouTube. This constitutes an exposure of sensitive local credentials and session tokens to the script and the underlyingyt-dlptool. - [PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection (Category 8) when processing transcripts and metadata from third-party videos.
- Ingestion points: Video transcripts and metadata downloaded from YouTube are interpolated into an LLM prompt template (
prompt_templates/clip_analysis_prompt.md). - Boundary markers: The prompt template relies on standard Markdown headers but lacks robust isolation or delimiters to separate untrusted transcript data from instructions, which could allow malicious content in a video transcript to influence the agent's behavior.
- Capability inventory: The script can read and write files, execute shell commands (via FFmpeg/Remotion), and perform network operations.
- Sanitization: While the code includes basic sanitization for output filenames and validates the JSON structure of the AI's response, it does not sanitize the input transcript content before processing, nor does it possess defenses against a transcript designed to manipulate the clipping logic.
Audit Metadata