# What is an LLM firewall?

An LLM firewall is a layer between users and a large language model (LLM) that inspects every prompt and response and stops the ones that break a policy. It is another name for input and output guardrails.

## How it works

An LLM firewall runs as a proxy in front of the model, or as an API the app calls around each model call.

- **Prompts** are checked before they reach the model, for [prompt injection](/glossary/prompt-injection), [jailbreaks](/glossary/jailbreak), personal data and unsafe requests.
- **Responses** are checked before they reach the user, for leaked data, unsafe content and fabricated facts.

A message that breaks a policy is blocked, rewritten or flagged. See [input and output guardrails](/glossary/input-and-output-guardrails).

## Limits of the analogy

Like a network firewall, it sits in the path of all traffic and applies rules the operator sets from outside the system it protects. The likeness ends there:

- **No fixed fields.** A network firewall matches addresses, ports and protocols exactly. A prompt is free text, so the verdict is a judgment of meaning by a [guard model](/glossary/guard-models), with false positives and misses.
- **No single perimeter.** Instructions also arrive in retrieved documents and tool results, and an agent's actions leave as tool calls, not as replies. A layer that reads only the chat misses both, so agents need [checks on tool calls](/glossary/tool-call-guardrails) too.

## With White Circle

[Control](/products/control) sits between your model and your users, analyzes every input and output in real time, and checks tool calls before they run.

