voice-agent
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: As an orchestration guide for voice agents, this skill describes an architecture that ingests untrusted user audio and text, which is an inherent attack surface for indirect prompt injection.
- Ingestion points: User audio streams and typed messages received via the
wss://agent.deepgram.com/v1/agent/converseWebSocket as described inSKILL.md. - Boundary markers: The skill mentions the use of
agent.think.promptfor system instructions, but does not provide specific guidance on sanitizing or delimiting transcription data before it is processed by the underlying LLM. - Capability inventory: The documented architecture supports
FunctionCallRequestfor client-side tool execution, binary audio generation (TTS), and mid-call updates to the prompt or model configuration. - Sanitization: Not explicitly detailed in the protocol reference; implementations rely on the safety guardrails of the chosen LLM provider (OpenAI, Anthropic, etc.).
- [EXTERNAL_DOWNLOADS]: The skill references official Deepgram repositories and SDKs for implementation and boilerplate code.
- Evidence: References official examples at
github.com/deepgram/examples,github.com/deepgram/recipes, andgithub.com/deepgram-starters. - Evidence: Lists official NPM packages including
@deepgram/agents,@deepgram/react,@deepgram/ui, and@deepgram/agents-widget. - Evidence: Mentions official SDK skills such as
deepgram-js-voice-agent,deepgram-python-voice-agent, anddeepgram-go-voice-agent.
Audit Metadata