# Continuous red teaming

Continuous red teaming is running attack suites against an AI system on every deploy and on a schedule, not once before launch. It tracks regressions, so a fixed hole that reopens is caught.

## Why once is not enough

A red team report describes the system on the day it was tested. After that, things move:

- **The model.** A new version, or a hosted model that its provider updates, behaves differently.
- **Your app.** A reworded system prompt, a new tool or a new data source can reopen a hole that was closed.
- **The attacks.** New [jailbreak](/glossary/jailbreak) and [prompt injection](/glossary/prompt-injection) techniques keep appearing.

## How it works

1. Keep a suite of attacks for each risk: jailbreaks, leaks, tool misuse.
2. Run it automatically on every deploy, and on a schedule between deploys.
3. Compare each run with the last. An attack that failed before and succeeds now is a regression.
4. Fix what is found, and add every new finding to the suite.

[How to red team an LLM](/glossary/how-to-red-team-an-llm) covers building the suite.

## With White Circle

[Red Team](/products/red-team) runs one-off campaigns or continuous testing, with regressions tracked over time. Findings can be patched through Control in one click, and we recommend testing on every deploy.

