voice-agents
Pass
Audited by Gen Agent Trust Hub on Aug 8, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill demonstrates an architecture that ingests untrusted user audio, transcribes it, and interpolates the resulting text directly into LLM prompts.
- Ingestion points: Transcribed text from Deepgram Nova-3 is assigned to
userTextin theprocessVoiceInputfunction inSKILL.md. - Boundary markers: The code does not implement delimiters or 'ignore' instructions for the interpolated user content.
- Capability inventory: The agent can generate LLM responses, synthesize speech via ElevenLabs, and control local audio playback.
- Sanitization: No evidence of input filtering or sanitization of the transcribed text before it enters the LLM context.
Audit Metadata