voice-ai-development
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio input that is transcribed into text and directly appended to the conversation history before being sent to an LLM. This creates a surface for indirect prompt injection if the audio contains malicious instructions.
- Ingestion points: The "VoiceAgent.process_audio_stream" method in "references/detailed-guide.md" ingests "audio_in" from potentially untrusted sources.
- Boundary markers: There are no explicit delimiters or system instructions used to isolate the transcribed user input from the rest of the prompt context.
- Capability inventory: The skill interacts with external LLM providers and supports tool execution for managing orders or other customer data.
- Sanitization: The provided code sketches do not demonstrate sanitization or validation of the resulting transcript before inclusion in the LLM context.
Audit Metadata