In plain English
A red team plays the bad guy on purpose. They try to break in, trick the system or make it misbehave, then report what worked so it can be fixed. With AI, that means trying to get a chatbot to say harmful things, leak private information or ignore its rules.
Example
Before launching a customer chatbot, a company has a small team spend two weeks trying to make it reveal other customers' orders and give dangerous advice. Each trick that works gets fixed before launch.
In short
Attack yourself first. Finding the weak spots before someone else does.
Why it matters
Red teaming is how companies find problems that normal testing misses, because it thinks like an attacker instead of a typical user. Major AI labs red team new models before release, and many AI safety rules and guidelines now expect it.
Common confusion
Red teaming vs. AI evals. AI evals measure how well a system does on a fixed set of tests, again and again. Red teaming is open-ended and creative, hunting for failures nobody has thought to test yet. What red teams find often becomes a new eval.