azure-ai-voicelive-ts
Pass
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a real-time voice-to-text and text-to-voice interface, which represents a surface for indirect prompt injection attacks where a user could manipulate the AI via spoken instructions.\n
- Ingestion points: Untrusted audio data is received from the user and processed in handlers like
onInputAudioTranscriptionCompletedandonResponseAudioDelta(SKILL.md).\n - Boundary markers: The provided code samples do not demonstrate the use of delimiters or specific safety instructions to isolate user-provided audio transcriptions from the system's core operating instructions.\n
- Capability inventory: The skill allows for function calling and session updates based on the AI's interpretation of these inputs, which could be abused if the agent is successfully manipulated.\n
- Sanitization: While structured function data is parsed using
JSON.parse, the natural language content derived from audio is not sanitized for prompt injection patterns.
Audit Metadata