Alignment & Safety

Red Teaming

Deliberately probing an AI system to find ways it can fail, be misused, or produce harmful output, before attackers do.

Red teaming is adversarial testing: people try hard to break a model, trick it into unsafe answers, or expose weaknesses, so those problems can be fixed before release. It is borrowed from security practice and has become standard for frontier models. The findings feed back into better guardrails and training, and the term red team refers to the people playing attacker.

Reference: Wikipedia: Red Teaming