Anthropic 在 AI 蒸馏战中将 Claude Token 盗窃率降低 99%
anthropic 通过新防御措施降低了claude推理token盗窃率90-99%，影响了zhipu和moonshot等中国公司的工业规模蒸馏攻击。
anthropic has successfully defended its claude models against industrial-scale distillation attacks, reducing the volume of extracted reasoning tokens by 90 to 99 percent. the attacks, carried out by networks of hundreds of thousands of fake accounts, were used by companies like zhipu and moonshot to copy claude's capabilities. zhipu allegedly used the stolen capabilities to train its glm-5.3 model, while moonshot secretly routed queries through claude. anthropic's new defenses include restricting access to claude's internal reasoning steps, running a classifier to detect extraction attempts, and automatically banning flagged accounts. tens of thousands of accounts have already been banned.