Leading Chinese AI laboratory Zhipu AI has been caught illegally using responses from Claude and GPT models to train its latest system, GLM-5.2. This statement was made at the Aspen Security Forum, where Anthropic's Head of National Security Policy, Tarun Chhabra, presented irrefutable evidence.
Chhabra noted that the current gap between the US and China in advanced AI models is only six to nine months. However, he stated that without unauthorized distillation—a method where a weaker model copies responses from a more powerful one—this gap would be as wide as a year and a half. He characterized Beijing's behavior as "hostile," emphasizing that competitors are essentially siphoning the most valuable intellectual property from American developments.
According to Anthropic, "virtually all" leading Chinese laboratories engage in distillation, using not only Claude but also models from other American developers. The company is actively blocking involved accounts—up to several million per week. As for Zhipu AI, their GLM-5.2 model was trained on responses from Claude and GPT, and DeepSeek is conducting a similar campaign.
Independent research firm Graphistry confirmed these findings. The Cohen's kappa coefficient between GLM-5.2 and GPT-5.5 responses was 0.8, and with Opus 4.8 it was 0.76. For comparison, the coefficient between GPT-5.5 and Opus 4.8 is 0.63. Zhipu AI has not yet provided comments on this matter.
Escalation of the US-China AI Conflict
Tensions in the AI sector are rising. In July, Chinese authorities declared Anthropic's Claude Code tool a "serious threat" to local users, after which Alibaba banned its employees from using it. In response, Washington launched the Gold Eagle initiative—a platform for sharing cybersecurity vulnerability data among leading US AI companies, which has already been joined by the Department of Defense and other agencies.
Treasury Secretary Scott Bessent stated that regulators will continue to use advanced AI models to protect against new threats. Earlier, the Trump administration blocked foreign access to Anthropic's powerful models—Fable 5 and Mythos 5—leading to a three-week shutdown of these neural networks for all users until access was restored.
My expert commentary: This situation demonstrates that the AI technology race is entering a phase of open confrontation. Distillation is not just a copyright violation but a direct threat to US national security. Investors should consider that such incidents will intensify regulatory pressure and technology export restrictions, which could significantly impact market dynamics and asset values in the sector.