Anthropic rolls out new protections, cutting data stolen in model-distillation attacks by up to 99%Anthropic 推出新防护措施，模型蒸馏攻击窃取数据量最高降低 99%

IT Home, Oct. 13 — Anthropic said on Monday local time that it has introduced several new protections designed to stop competitors from replicating the capabilities of its Claude AI models. The measures have already reduced the amount of data stolen in model-distillation attacks by about 90% to 99%. Model distillation has become a major concern for leading AI companies: Anthropic, OpenAI, Google and others believe it lets competitors replicate a model's frontier capabilities faster and more cheaply. Anthropic says the protections, which limit outside access to Claude's reasoning process, have sharply cut information-extraction behaviour: a distillation attacker can now extract only one-tenth to one-hundredth of the reasoning tokens they could before. The California-based company says the new protections make it harder to repeatedly probe Claude's protected reasoning chain via proxy networks, block extraction attempts and hamper large-scale distillation campaigns. Anthropic describes the safety system's robustness as roughly comparable to the protections that stop Claude from helping to make biological weapons, with a false-positive rate below 0.1%. Anthropic has banned tens of thousands of accounts linked to distillation activity and, citing AI start-up Zhipu AI as an example, said the company tried to distil capabilities from Claude while developing its GLM-5.3 model. Anthropic warns that such behaviour carries security and national-security risks.