Answer
A dedicated team acts like an attacker, testing the AI system with tricky, unusual, or adversarial inputs to uncover vulnerabilities that need to be fixed before deployment.
Real-world example A company runs an AI red teaming exercise, having a dedicated team try to trick their chatbot into giving harmful responses before it is released publicly.