Red Teaming

Deliberately attempting to elicit harmful, incorrect, or unintended behavior from an AI system — an adversarial evaluation practice borrowed from cybersecurity.