Reference · Glossary

Red teaming

Deliberately **attacking your own AI system** to find failures — unsafe outputs, leaks, bad tool calls — before users do.

When to use

Before launching customer-facing bots, agents with tools, or apps handling sensitive data.

When not to

A solo weekend hackathon with no users — basic prompt testing may be enough initially.

Example

A red team tries 200 adversarial prompts (injection, bias bait, PII extraction) and files bugs ranked by severity.