New Defenses Reduce AI Model Distillation Rates by 90%

A new defense mechanism reduces the success rate of illicit model distillation by more than 90%. This approach prevents attackers from extracting frontier model capabilities through targeted query patterns.