voice-ai-integration

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.65). The skill’s runtime flow ingests outsider-authored free text via the Hume EVI transcript/messages that are produced from the user’s spoken input (and thus can include arbitrary user-provided text) into the agent/LLM context through the “real-time transcript display” / “conversation turns table” path; this is not operating-user-authored content.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill opens runtime WebSocket connections to the Hume EVI endpoint (wss://api.hume.ai/v0/evi/chat) — used by the client and Twilio outbound calls — which executes remote Hume model inference to generate the AI's behavior, making it a runtime external dependency that controls agent output.