Anthropic has implemented new security measures to prevent the distillation of its AI models, successfully reducing the volume of data extracted from Claude by 10 to 100 times. These protections specifically target the extraction of reasoning tokens, which competitors previously used to replicate the capabilities of advanced systems at a lower cost.
The company achieved these results by restricting the use of proxy networks that facilitated large-scale data harvesting. Anthropic stated that the reliability of these new defense mechanisms is comparable to its existing safety protocols designed to prevent the creation of biological weapons, with a false-positive rate of less than 0.1 percent.
As part of its enforcement efforts, Anthropic identified and blocked several tens of thousands of user accounts linked to unauthorized data extraction attempts. The company specifically accused the Chinese startup Z.ai, also known as Zhipu, of attempting to intercept Claude's outputs during the development of its GLM-5.3 model, citing national security concerns.
