ElevenAgents Guardrails

Guardrails 2.0 is the safety control layer inside ElevenLabs' ElevenAgents platform, built for teams putting voice agents on real calls in support, sales, and internal workflows. It isn't a standalone moderation API. Instead it wraps each live conversation in layered checks: the Focus Guardrail reinforces your system prompt so the agent stays on task through long calls, Manipulation Guardrails scan caller input for prompt injection and instruction-override attempts and can end a risky call, and Content Guardrails screen every agent reply against tunable sensitive-content categories. Custom Guardrails stretch the policy layer into your own domain: you describe rules in natural language and a lightweight model returns block or allow for each response, evaluated independently and in parallel with generation. The part that separates this from generic moderation services is enforcement. On a trigger you choose the exit: end the call, transfer to a different agent, escalate to a human, or retry with corrective instructions. Execution modes trade strictness for latency — run checks alongside the response for near-zero delay, or hold the reply until it clears. Content sensitivity is tunable per category, guardrails toggle individually per agent, and every trigger lands in conversation analytics with which guardrail fired and what action followed. Enterprise clients add redaction of transcripts, recordings, and webhook payloads — placeholders in text, bleeps in audio — with granularity down to individual entity types, plus Zero Retention Mode for stricter compliance. The full Guardrails 2.0 suite is available in alpha, configured in the Security tab or via the API, and it sits beside newer Eleven v4 speech models that add emotional nuance and faster responses across the platform.