The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides Python code that uses subprocess.run with list-based arguments to interact with ffmpeg and ffprobe. This is a secure implementation that prevents shell injection attacks.
[PROMPT_INJECTION]: The skill creates an indirect prompt injection surface (Category 8) by transcribing audio from untrusted sources and feeding the resulting text into downstream LLM agents. 1. Ingestion points: Audio data enters the system through the validate_audio_file and preprocess_audio functions. 2. Boundary markers: The transcripts use structured JSON and speaker tags, but lack explicit delimiters or instructions to prevent the downstream agent from executing commands embedded in the text. 3. Capability inventory: The pipeline has access to system command execution via subprocess and network operations via httpx. 4. Sanitization: The normalize_transcript function performs basic text formatting cleanup but does not filter content for adversarial instructions.

engineering-voice-ai-integration-engineer