Techniques & safeguards
Red teaming and adversarial testing
Standards, techniques & other
Deliberately attacking a model to find failures: jailbreaks, harmful outputs, data leakage, bias or security holes. Adversarial testing is mandatory for GPAI models with systemic risk.