split-audio-by-words
Pass
Audited by Gen Agent Trust Hub on Aug 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes ffmpeg shell commands to slice audio files and apply fade effects. While this is the intended purpose, it involves direct interaction with the system shell.
- Evidence: The workflow utilizes
ffmpeg -y -ss "$snapped_start" -to "$snapped_end" -i "$audio" -af ...for extracting scene splits. - [EXTERNAL_DOWNLOADS]: Audio data is transmitted to external AI providers (Groq, OpenAI, and Fal) to perform word-level transcription.
- Evidence: The skill calls Groq Whisper, OpenAI Whisper, and the external
atoms/source/transcribe-audio-falto process the<audio>input. - [PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection by transcribing untrusted audio content into a structured manifest and transcript files.
- Ingestion points: Untrusted audio files are ingested in Step 1 and converted to text via Whisper transcription.
- Boundary markers: No delimiters or "ignore embedded instructions" warnings are applied to the transcribed text stored in
manifest.jsonorwords.json. - Capability inventory: The skill performs file system writes and executes shell commands via
ffmpeg. - Sanitization: No sanitization, escaping, or filtering of the transcribed content is mentioned before it is stored or displayed in the verification report.
Audit Metadata