One of China's leading artificial intelligence companies, Zhipu, trained its GLM-5.2 model using responses from Claude and GPT. This was stated by Tarun Chhabra, head of national security policy at Anthropic, during the Aspen Security Forum, as reported by the South China Morning Post.
According to him, the U.S. is ahead of China in advanced AI models by six to nine months. Without unauthorized distillation, the gap would be even larger—up to a year and a half. Chhabra described this behavior from Beijing as "hostile," emphasizing that competitors are extracting the most valuable intellectual property from foreign models.
"Almost all" leading Chinese labs are distilling, not just Claude, but also models from other American developers. Anthropic is actively blocking involved accounts—up to several million per week.
Chhabra also reported new evidence regarding Zhipu AI. Its latest model, GLM-5.2, was trained on responses from Claude and GPT. He added that DeepSeek continues a similar "hostile campaign."
Similar findings were previously published by the research company Graphistry. The Cohen's kappa coefficient indicated that the responses of GLM-5.2 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 0.63. Zhipu AI has neither confirmed nor denied this assessment.
U.S. Expands Measures Against China in AI
In July, Chinese authorities labeled Anthropic's Claude Code as a "serious threat" to Chinese users. Following this, Alibaba prohibited its employees from using the tool for security reasons.
Meanwhile, Washington is tightening its AI policy regarding Beijing. The White House previously launched the Gold Eagle initiative—a platform where leading U.S. AI companies share data on cybersecurity vulnerabilities. Several American agencies, including the Department of Defense, are already collaborating with it.
Secretary of the Treasury Scott Bessent stated that the regulator will continue to use advanced AI models to "stay ahead of adversaries and protect Americans from new threats."
This follows a decision by the Trump administration to restrict foreign access to Anthropic's most powerful models—Fable 5 and Mythos 5. As a result, the company temporarily disabled both neural networks for all users for nearly three weeks. Access was later restored.
In February, Anthropic accused DeepSeek, Moonshot AI, and MiniMax of distilling Claude through thousands of fake accounts. The company later described the campaign associated with Alibaba as the largest of the known attacks of this kind.