Leading Chinese AI developer Zhipu has found itself at the center of a major scandal. It has emerged that its latest model, GLM-5.2, was trained on responses from American systems Claude and GPT. This statement was made by Tarun Chhabra, head of national security policy at Anthropic, at the Aspen Security Forum.
According to the data presented, the United States maintains its lead in developing advanced AI models, outpacing China by six to nine months. However, in the expert's opinion, without the practice of unauthorized distillation, this gap could have reached a year and a half. Distillation refers to a training method where a less powerful model copies the behavior of a larger, more complex system. Beijing's actions were characterized as "hostile" and aimed at extracting valuable intellectual property.
Notably, "virtually all" leading Chinese laboratories engage in distillation. They are not limited to cloning Claude but also actively use models from other American developers. In response, Anthropic is forced to block up to several million accounts weekly to prevent unauthorized access. It was separately noted that DeepSeek continues a similar campaign, which only confirms the systemic nature of the problem.
Independent research firm Graphistry has also confirmed these findings. According to their analysis, GLM-5.2's responses match GPT-5.5 with a Cohen's kappa coefficient of 0.8, and Opus 4.8 with 0.76. For comparison, the coefficient between GPT-5.5 and Opus 4.8 themselves is only 0.63. Zhipu AI, for its part, has neither confirmed nor denied these allegations.
Escalation of the Technological Standoff
The situation is worsening amid the tightening of U.S. policies toward Chinese AI companies. Earlier, Chinese authorities recognized Anthropic's Claude Code tool as a "serious threat" to its users, leading to a ban on its use by Alibaba employees. In response, Washington launched the Gold Eagle initiative—a platform for sharing cybersecurity vulnerability data among leading American AI companies. U.S. Treasury Secretary Scott Bessent stated that regulators will actively use advanced models to stay ahead of adversaries.
The culmination came with the Trump administration's decision to block foreign access to Anthropic's most powerful models—Fable 5 and Mythos 5. This led to nearly three weeks of disconnection for all users, after which access was restored. Recall that back in February, Anthropic accused DeepSeek, Moonshot AI, and MiniMax of large-scale distillation through thousands of fake accounts, and later described the campaign linked to Alibaba as the largest known attack of its kind.
My analysis: This situation demonstrates not just technological piracy but a systemic conflict of interests at the level of national security. Chinese companies, in their bid to close the gap, are resorting to risky and essentially illegal methods. For investors, this is a troubling signal: dependence on "foreign" code makes such models vulnerable and unsustainable in the long term. The AI market is entering a phase of fierce competition, where intellectual property will become the main battlefield.