# What are AI guardrails?

AI guardrails are checks that run outside a model on its inputs, outputs and tool calls. When something breaks a policy, the guardrail acts on it, for example by blocking, rewriting or flagging it.

## How they work

A guardrail sits in the request path, between the application and the model. It judges what passes against each [policy](/glossary/guardrail-policy) and takes the [action](/glossary/guardrail-actions) of any policy that is broken. The check can be a pattern match, a classifier, or a model trained to judge text (a [guard model](/glossary/guard-models)).

There are three places to check:

- **Inputs**, before the model call: unsafe requests, prompt injection, personal data.
- **Outputs**, before the user sees them: leaks, fabricated facts, unsafe content.
- **Tool calls**, before they run: an agent's commands, code and API calls.

## Why prompts alone are not enough

A system prompt asks the model to behave. The same model also reads the user's message and every document it is given, and a [jailbreak](/glossary/jailbreak) or an injected instruction can talk it out of its rules. Nothing then stops the reply.

A guardrail does not rely on the model's cooperation. It is a separate check that applies the same policy to every request and enforces the result. See [system prompt vs guardrails](/glossary/system-prompt-vs-guardrails).

## With White Circle

[Control](/products/control) sits between your model and your users and analyzes every input and output in real time. It blocks, rewrites or flags anything that breaks your policies, and checks tool calls before they run.

