Anthropic's New Shields Cut AI Model Theft Data by Up to 99%

Model distillation lets smaller models learn from bigger ones to copy their abilities. Top AI companies worry rivals could use this to rebuild frontier models faster and at lower cost.

Anthropic said its new protections have cut the data stolen in distillation attacks by about 90% to 99%. Attackers can now only get one-tenth to one-hundredth of Claude's reasoning tokens compared to before, and it is harder to use proxy networks to probe the protected reasoning chains.

Anthropic said this system is about as strong as the one that stops Claude from helping make biological weapons, with a false alarm rate below 0.1%. The company has also banned tens of thousands of accounts linked to distillation.

Anthropic pointed to Chinese AI startup Zhipu AI, saying it tried to copy Claude's abilities while building the GLM-5.3 model. The company warned such moves bring security and national security risks.