Anthropic reveals theft of AI technologies: Chinese Zhipu accused of distilling Claude and GPT
Serious accusations against China's AI sector were voiced at the Aspen Security Forum. Tarun Chhabra, head of national security policy at Anthropic, stated that one of China's leading technology corporations, Zhipu, trained its GLM-5.2 model on the responses of Claude and GPT models. This is a direct violation of intellectual property norms.
According to the expert's estimates, the current gap between the US and China in advanced AI models is only six to nine months. However, without unauthorized distillation—a method where a less powerful model learns from the responses of a larger one—this gap could have reached one and a half years. Chhabra called such actions by Beijing "hostile" and emphasized that competitors are extracting the most valuable intellectual property from others' models.
"Virtually all" leading Chinese laboratories are engaging in distillation, targeting not only Claude but also models from other American developers. Anthropic is actively blocking accounts involved in this process—up to several million per week. Separately, Chhabra reported new evidence regarding Zhipu AI: the GLM-5.2 model was trained on responses from Claude and GPT. DeepSeek continues a similar "hostile campaign."
Research firm Graphistry confirmed these findings. According to Cohen's kappa coefficient, GLM-5.2 responses matched GPT-5.5 at 0.8 and Opus 4.8 at 0.76. For comparison, the score between GPT-5.5 and Opus 4.8 was only 0.63. Zhipu AI neither confirmed nor denied this assessment.
US Tightens AI Policy Against China
In July, Chinese authorities called Anthropic's Claude Code a "serious threat" to Chinese users, after which Alibaba banned its employees from using this tool. Meanwhile, Washington is intensifying measures: the White House launched the Gold Eagle initiative—a platform where leading US AI companies share data on cybersecurity vulnerabilities. The Department of Defense and other agencies are already working with it.
US Treasury Secretary Scott Bessent stated that the regulator will continue to use advanced AI models to "stay ahead of adversaries and protect Americans from new threats." Earlier, the Trump administration blocked foreign access to Anthropic's most powerful models—Fable 5 and Mythos 5—leading to the temporary shutdown of these neural networks for all users.
Recall that in February, Anthropic already accused DeepSeek, Moonshot AI, and MiniMax of distilling Claude through thousands of fake accounts. Later, the company called the Alibaba-linked campaign the largest known attack of its kind.
My analysis: This escalation is not just a corporate dispute but a reflection of the geopolitical struggle for AI leadership. Distillation undermines the fundamental principles of innovation, making Chinese models cheap copies rather than independent developments. If the US fails to protect its intellectual property, the technology gap could shrink faster than expected, altering the global balance of power in this critical industry.