Systematic probing of an AI model with adversarial inputs to discover safety, security, and alignment weaknesses before release.