ElevenAgents — agentic threat model

Integrates with external foundation models (Gemini, Claude, OpenAI, or custom LLMs) and uses custom STT/TTS models. Key threats include voice-based prompt injection (adversarial audio), model misalignment, and potential model stealing of the fine-tuned ASR/TTS assets.

Supports knowledge base integration via document uploads. This introduces risks of data/knowledge-base poisoning (uploading malicious or misleading documents to alter agent behavior) and unauthorized data exfiltration of sensitive uploaded documentation through conversational extraction.

Not certain from the listing — the specific orchestration framework and turn-taking logic are proprietary. Potential threats include framework vulnerabilities in the custom turn-taking and interruption-handling logic, as well as insecure integration with the underlying LLM APIs.

Not certain from the listing — hosting, API gateway security, and sandboxing mechanisms are not detailed. Threats include infrastructure compromise of the low-latency streaming servers, API key exposure, and denial-of-service attacks targeting the real-time voice processing pipeline.

Not certain from the listing — no built-in guardrails, evaluation metrics, or observability tools are mentioned. Threats include conversational drift, lack of audit trails for voice interactions, and blind spots in detecting malicious inputs translated through the STT engine.

Not certain from the listing — compliance certifications (such as SOC2, GDPR, or HIPAA for voice recordings) and access control policies are not specified. Threats include unauthorized access to call logs/recordings and lack of robust authentication for API-driven agent deployments.

Mentions multi-agent engagement to resolve issues. This introduces risks of agent-to-agent trust abuse, cascading failures if one agent in the pipeline is compromised or returns unexpected outputs, and coordination failures during complex customer support handoffs.