A team plays the attacker: bypassing guardrails, obtaining forbidden content, extracting data. For AI applications, red teaming combines hand-written scenarios and automatically generated attacks.
It is the AI equivalent of penetration testing in classic security.