deepgram-python-voice-agent
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of interactive voice agents that ingest real-time audio and user-supplied text messages. These inputs are processed by an LLM to generate responses or trigger functions, creating a surface where malicious instructions could be embedded in the data stream.
- Ingestion points: Untrusted data enters the agent context via audio frames in
agent.send_media()(SKILL.md) and text content inagent.send_inject_user_message()(SKILL.md). - Boundary markers: No explicit delimiters or "ignore embedded instructions" warnings are used when configuring the LLM prompt or processing turn-based messages.
- Capability inventory: The agent has the ability to execute tool/function calls (
FunctionCallRequest) and maintain stateful WebSocket connections. - Sanitization: The provided code snippets do not include input validation, escaping, or filtering for the audio or text content being processed.
- [EXTERNAL_DOWNLOADS]: The documentation suggests installing central product skills via
npx skills add deepgram/skills. This command downloads and executes content from the vendor's official repository. This is an expected installation pattern for the Deepgram ecosystem.
Audit Metadata