deepgram-java-voice-agent
Pass
Audited by Gen Agent Trust Hub on Jul 16, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill demonstrates secure credential management by using environment variables (
System.getenv("DEEPGRAM_API_KEY")) instead of hardcoding secrets. - [EXTERNAL_DOWNLOADS]: The skill suggests installing additional resources using
npx skills add deepgram/skills. This command targets the official repository of the skill author and uses a standard installation tool. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted user audio and text input, creating a potential surface for indirect injection.
- Ingestion points: User data enters through the
V1WebSocketClientaudio and text streams inSKILL.md. - Boundary markers: The provided Java code examples do not include explicit boundary markers or instructions to the model to ignore embedded commands in the user audio.
- Capability inventory: The SDK supports dynamic prompt updates (
sendUpdatePrompt) and message injection (sendInjectUserMessage), which are controlled via the WebSocket session inSKILL.md. - Sanitization: The implementation examples do not demonstrate sanitization or validation of the ingested user content.
- [SAFE]: No evidence of obfuscation, persistence mechanisms, or unauthorized privilege escalation was detected.
Audit Metadata