What is AI control?

AI control is enforcing what an AI system may say and do at run time, whatever the model itself intends. Policies are checked outside the model on every input, output and tool call.

How it works

Control assumes the model cannot be fully trusted: it may be wrong, jailbroken or injected. So the rules are enforced by a separate layer of guardrails, not by the model's own judgment.

Inputs are checked before the model reads them, outputs before the user sees them, and tool calls before they run. Each check compares the event with a set of policies and acts: allow, block, rewrite, hold for approval, end the session.

In safety research the term has a stricter sense: measures that keep a system safe even if the model is actively working against them.

Control and alignment

Alignment shapes what a model tends to do, through training. Control limits what it can do, whether or not the training worked. Training lowers the failure rate but not to zero, and an agent's failure is an action: a payment, a deleted file. Control is what still holds when training fails.

A system prompt is not control: it asks, and the model can be talked out of it. See system prompt vs guardrails.

With White Circle

This is what our platform does. Control enforces policies written in plain language on every input, output and tool call in real time: it blocks, rewrites or flags what breaks them.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL