IT Home On October 13, Anthropic said on Monday local time that it has introduced several new security measures to prevent competitors from replicating the Claude AI model. These measures have reduced the amount of data stolen in model distillation attacks by approximately 90% to 99%.
Model distillation has become a concern for top AI companies. Companies such as Anthropic, OpenAI, and Google believe that this approach allows competitors to replicate the cutting-edge capabilities of models at lower costs and faster speeds.
According to Anthropic, this protection mechanism used to limit external access to Claude’s reasoning process has significantly reduced information extraction behavior. Now, the number of reasoning tokens that “distill attackers can extract is only one-tenth to one-hundredth of what it was before”.
IT Home noticed that the company headquartered in California stated that the new protection increased the difficulty of repeatedly probing the Claude protected reasoning chain via proxy networks, intercepted information extraction attempts, and hindered large-scale distillation operations.
According to Anthropic, the robustness of this security system is “roughly equivalent to a protection system that prevents Claude from assisting in the creation of biological weapons”, and the false positive rate is less than 0.1%.
Anthropic has banned tens of thousands of accounts related to distillation activities. Taking the AI startup Zhipu AI as an example, it said that during the development of the GLM-5.3 model, the company attempted to extract the model’s capabilities from Claude distillation. Anthropic warns that such actions pose risks to security and national security.