Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillation

If you use an AI assistant, its core reasoning performance may improve and its costs could decrease.

Anthropic announced it has developed defenses to combat the industrial-scale distillation of its AI model's reasoning capabilities, reducing the volume of extracted internal chain-of-thought steps by 90% to 99%. These attacks involve the systematic copying of an AI model's functionalities to train smaller, proprietary models. Anthropic identified Chinese companies Zhipu and Moonshot as engaged in these practices; Zhipu allegedly used stolen capabilities for its GLM-5.3 model, while Moonshot reportedly routed queries through Claude. The company claims a sharp drop in successful distillation attempts after implementing its new defenses, which include restricting access to reasoning steps, deploying a classifier to detect extraction attempts, and banning offending accounts, with tens of thousands already banned. Anthropic is urging other AI developers to adopt similar measures, acknowledging that legitimate users may also be inadvertently affected.

The success of Anthropic's defenses against AI model distillation signifies a critical hardening of AI intellectual property. By drastically reducing the ability of external actors to siphon proprietary reasoning techniques, Anthropic is aiming to preserve the competitive edge of its large language models. This move impacts the broader AI ecosystem by potentially raising the barrier to entry for companies seeking to quickly replicate advanced AI capabilities without significant R&D investment, forcing them to develop original architectures or accept higher risks of detection and account suspension.

The implications extend to national security and the global AI race, as Anthropic highlights the potential diversion of stolen AI capabilities into military and surveillance systems. The company's call for industry-wide adoption of similar protections suggests a nascent arms race in AI security, where AI developers must balance robust defense mechanisms against the risk of alienating or hindering legitimate research and application. Future developments will likely involve a cat-and-mouse game between distillation techniques and increasingly sophisticated detection and prevention methods.