voice-ai-development

Fail

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: HIGHPROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection (HIGH): The skill is specifically designed to ingest untrusted external data (user voice audio and transcripts) and pass it to LLMs that possess side-effect capabilities (tool/function calling).
  • Ingestion points: Data enters through wss://api.openai.com/v1/realtime (OpenAI), request.json in the /vapi/webhook route (Vapi), and audio_stream parameters (Deepgram/ElevenLabs).
  • Boundary markers: The provided code snippets do not include system prompt delimiters or 'ignore embedded instructions' warnings for the audio/text being processed.
  • Capability inventory: The templates demonstrate tool-calling capabilities such as get_weather (OpenAI Realtime) and check_order (Vapi webhook). Successful injection via user speech could lead to unauthorized tool execution.
  • Sanitization: No sanitization or validation logic is shown for the transcripts or parameters extracted from the voice interactions before they are used in business logic or tool calls.
  • Data Exposure & Exfiltration (LOW): The code snippets include placeholders for API keys (sk-..., api_key="..."). While these are not actual credentials, users must ensure they manage their secrets via environment variables rather than hardcoding them as shown in the examples.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 16, 2026, 07:49 AM