Safety Control Layer

Guardrails 2.0 for ElevenAgents is a new control layer that helps voice agents maintain safety, brand consistency, and regulatory compliance during conversations. Since system prompts alone are insufficient, it is designed around a multi-layer defense structure that independently checks both user input and agent responses.

Three built-in protections are provided.

• Focus Guardrail: Reinforces the system prompt so responses stay on-goal and within guidelines
• Manipulation Guardrails: Detects and blocks prompt injection or instruction-override attempts
• Content Guardrails: Screens for sensitive or dangerous content across multiple categories, with finely adjustable thresholds

Custom Guardrails can be added on top, automatically applying domain policies defined in natural language to every call. A lightweight model evaluates each response separately to make an allow / block decision, operating in parallel with response generation to minimize latency.

Operation can also be finely tuned. You can trade off strictness and latency by running guardrails alongside the response for near-zero delay, or by fully completing the check before releasing the response. When a violation occurs, you can choose an exit strategy such as ending the conversation, switching to a different agent, escalating to a human, or retrying with corrective instructions attached.

It also offers Content sensitivity levels, the ability to enable/disable individual guardrails, different settings per agent, and complete visibility through logs of triggers and actions taken. After a conversation ends, sensitive information in transcripts, recordings, and webhook payloads can be masked via Conversation History Redaction, which works alongside Zero Retention Mode to meet stronger compliance requirements.

These features are aimed at enterprise customers and are currently offered in alpha. They can be turned on from the Security tab or configured via API, and more broadly, they also underpin support for AIUC-1 certification and agent insurance programs.