Red teaming

Red teaming is the practice of deliberately attacking an AI system, often with specialized testers, to uncover harmful outputs, security weaknesses and failure modes before real users or attackers find them.

Red teaming sits in the Evaluation & safety part of the Agentik {OS} glossary, which defines the words used to build and run AI agent systems.

Also called AI red teaming, Adversarial testing.