elevenlabs-agents
Pass
Audited by Gen Agent Trust Hub on Jul 11, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill outlines an architecture for voice agents that ingest untrusted data from external callers and knowledge base documents, creating a potential surface for indirect prompt injection.
- Ingestion points: The agent processes real-time audio and text from external callers and content from uploaded documents (PDF, URL imports, etc.) as described in the "Knowledge base" and "Channels and SDKs" sections of SKILL.md.
- Boundary markers: SKILL.md identifies the use of a "trust_context" configuration and recommends "scripting the edges" within the system prompt to handle ambiguous or out-of-scope user inputs.
- Capability inventory: The agent is designed to perform server-side webhooks, client-side JavaScript execution, and session management tasks according to the "Tools and function calling" section.
- Sanitization: The instructions advocate for enforcing brevity, banning formatting in responses to protect text-to-speech output, and validating tool parameters using regex or LLM-based semantic checks.
Audit Metadata