verysmallwoods-video
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill is designed to download audio (.mp3) and subtitle files (.srt, .vtt) from the ListenHub API (api.marswave.ai) as part of its core text-to-speech workflow. It also bridges to Groq or OpenAI for transcription services in fallback scenarios.
- [COMMAND_EXECUTION]: The srt_helper.py script executes the ffmpeg command via subprocess.run to perform necessary audio adjustments, including inserting silent gaps and concatenating media segments. This operation is considered safe and standard for the skill's declared purpose.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided narration text which is eventually passed to an LLM for subtitle correction. This creates a surface for indirect prompt injection. Ingestion points: Narrations are ingested from a text file (narration.txt) as described in the workflow. Boundary markers: The srt_helper.py correct subcommand lacks explicit boundary markers or instructions to isolate potentially malicious commands within the subtitles. Capability inventory: The skill possesses capabilities for file system writes, network requests (via curl and urllib), and local command execution (via ffmpeg). Sanitization: Input text is processed without explicit sanitization or filtering before being sent to the LLM.
Audit Metadata