Anthropic implements security measures to prevent AI model distillation

Anthropic has implemented new security measures to protect its Claude AI models from “distillation” attacks, which allow competitors to replicate advanced AI functionalities at lower costs and faster speeds. The company reports that these new protections have reduced the amount of data extracted through such attacks by approximately 90% to 99%.

The security mechanisms specifically limit access to Claude’s reasoning processes, making it significantly harder for rivals to extract information for developing their own systems. Anthropic stated that the number of reasoning-related tokens extracted by attackers is now “10 to 100 times smaller than before.”

As part of these efforts, Anthropic has blocked tens of thousands of accounts associated with distillation campaigns. The company cited the Chinese startup Zhipu AI (also known as Z.ai) as an example of an entity that attempted to copy Claude’s capabilities during the development of its GLM-5 model. Anthropic warned that such practices pose risks to safety and national security amidst the intensifying technological competition between the United States and China.