twilio-voice-conversation-relay

Pass

Audited by Gen Agent Trust Hub on Jul 6, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill processes voicePrompt data, which contains ASR-transcribed speech from external callers. This untrusted data is intended to be passed to an LLM, creating a surface for indirect prompt injection.
  • Ingestion points: SKILL.md contains code examples (Python and Node.js) that extract voicePrompt from incoming WebSocket messages.
  • Boundary markers: While the code examples do not include boundary markers, the documentation includes a 'Security' advisory explicitly warning developers to isolate this input within a structured system prompt.
  • Capability inventory: The skill facilitates bidirectional audio and LLM integration. Exploitation could lead to the LLM disclosing system instructions or speaking unintended content to the caller.
  • Sanitization: The provided code snippets do not perform sanitization; however, the documentation advises implementing topic boundaries and output filtering.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 6, 2026, 07:49 PM
Security Audit — agent-trust-hub — twilio-voice-conversation-relay