Anthropic introduces enhanced security measures to counter model distillation attacks

Anthropic, a leading AI company, has unveiled new security measures to prevent competitors from replicating its Claude AI model's capabilities through model distillation attacks. These measures have reduced the amount of data stolen during such attacks by up to 99%. Model distillation, a method where attackers extract knowledge from a model, has become a major concern for companies like Anthropic, OpenAI, and Google, as it allows rivals to replicate advanced capabilities at lower cost and faster speed. The new security system significantly limits the ability of external actors to access Claude's internal reasoning processes, reducing the number of tokens extracted by attackers to as little as one-tenth or one-hundredth of previous levels. Anthropic claims the system is robust enough to prevent Claude from aiding in the development of biological weapons, with a false positive rate below 0.1%. The company has also blocked tens of thousands of accounts linked to distillation activities, citing an example involving Zhipu AI, which attempted to extract model capabilities during the development of its GLM-5.3 model. Anthropic warns that such activities pose risks to both security and national security, emphasizing the need for stronger protections against intellectual property theft in the AI industry.