ElevenLabs Rolls Out Guardrails 2.0 to Stop AI Agents From Going Off Script

ElevenLabs has launched Guardrails 2.0, a new safety system built into its ElevenAgents platform that gives businesses stronger control over how their AI voice agents behave during real conversations.

The update introduces a three-layer protection system designed to keep agents on topic, block manipulation attempts, and screen every response before it reaches the user. All of this happens in real time, with almost no added delay to conversations.

The first layer reinforces the agent's original instructions throughout an entire conversation, preventing it from drifting off course. The second layer catches and blocks adversarial inputs, such as attempts to override the agent's programming, and can automatically end risky conversations. The third layer checks every response against set policies before it is delivered.

One of the most useful additions is the custom guardrails feature, which lets businesses write simple rules in plain language. For example, a company can instruct the agent to never give financial advice or avoid discussing medical treatments. These rules are enforced automatically without any coding.

Businesses can also choose how strictly each guardrail runs, decide what happens when a rule is triggered, and even escalate conversations to a human agent when needed.

For enterprise users, the platform also offers automatic redaction of sensitive information from transcripts and recordings, along with tools for pre-launch testing and ongoing monitoring.

Guardrails 2.0 is live now inside ElevenAgents and can be enabled through the Security tab in the agent settings.