Anthropic launches new protective shield model distillation and calls out China’s “Wisdom Spectrum” for plagiarism

Anthropic, an American artificial intelligence (AI) company, said on the 12th that in order to prevent competitors from copying the capabilities of its AI model Claude through "model distillation" technology, it has introduced new measures to significantly reduce the amount of data that "distillation" can capture by 90% to 99%. Anthropic also named the Chinese AI startup Zhipu, saying that it had tried to develop its own model through "distillation".

Reuters reported on the 12th that Anthropic pointed out that the new protection measures restrict external access to Claude's reasoning content, significantly reducing the amount of data being captured. The number of reasoning tokens that distillers can extract is now "reduced to 1/10 to 1/100 of the previous level." The company is also making it more difficult to use a "proxy network" to repeatedly probe the content of Claude's protected inferences, blocking data extraction and making large-scale model distillation more difficult.

As the AI ​​competition between the United States and China intensifies, “model distillation” has become a major concern for leading AI industry players. Companies such as Anthropic, OpenAI and Google believe that competitors can use this approach to replicate the capabilities of advanced AI models at a lower cost and faster speed.

Anthropic stated that the safety level of this protection mechanism is approximately equivalent to "a safety system that prevents Claude from being used to assist in the production of biological weapons" and that the "misconception rate is less than 0.1%."

The company has blocked tens of thousands of accounts related to model distillation activities, and named the Chinese AI startup Z.ai, saying that it tried to obtain relevant capabilities from Claude "Distillation" during the development of the GLM-5.3 model. Anthropic warned that such behavior could pose safety and national security risks. Zhipu has not yet responded to this matter.