wake-word-detection
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFE
Full Analysis
- [DATA_EXFILTRATION]: The skill implements logic for continuous microphone monitoring, which involves accessing sensitive audio data. Security is maintained by strictly processing all audio locally, avoiding any network transmission of voice data, and utilizing circular buffers that are immediately cleared after use to prevent data persistence.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests raw audio data via the sounddevice library, representing an external input surface. The risk is minimized by the skill's specific design focus on local keyword spotting, which acts as a narrow filter to trigger actions rather than passing unsanitized content directly to the agent's prompt.
- [EXTERNAL_DOWNLOADS]: All identified dependencies (openwakeword, numpy, sounddevice, onnxruntime, webrtcvad) are standard, well-reputed libraries for signal processing and machine learning. No execution of scripts from untrusted remote sources or suspicious download-then-execute patterns were found.
- [DYNAMIC_EXECUTION]: The skill utilizes ONNX Runtime for keyword detection. The implementation uses standard session options and local model loading, avoiding unsafe dynamic code generation, runtime compilation of untrusted input, or insecure deserialization patterns.
Audit Metadata