Leading Chinese AI developer Zhipu AI has come under fire after it was revealed that their latest model, GLM-5.2, was trained on responses from Claude (Anthropic) and GPT (OpenAI) systems. This statement was made at the Aspen Security Forum by a senior representative from Anthropic.
According to expert estimates, China's current lag behind the United States in advanced AI models ranges from six to nine months. However, without systematic and unauthorized distillation — a method where a less powerful model "siphons" knowledge from a larger one — this gap could have reached one and a half years. Such actions by Beijing are characterized as openly hostile, aimed at appropriating the most valuable intellectual property.
It is important to emphasize that distillation is not an isolated incident but a widespread practice covering virtually all leading Chinese laboratories. They are not limited to Claude models but actively target developments from other American companies as well. In response, Anthropic has been forced to block up to several million accounts per week involved in such operations.
Special emphasis is placed on Zhipu AI and their GLM-5.2 model. Research firm Graphistry conducted an independent analysis and identified an anomalously high correlation between GLM-5.2's responses and the outputs of GPT-5.5 and Opus 4.8. Cohen's kappa coefficient was 0.8 and 0.76, respectively. For comparison, the agreement between GPT-5.5 and Opus 4.8 themselves is estimated at only 0.63. This serves as compelling evidence of direct borrowing. Zhipu AI has not yet provided any refutations or confirmations of this data.
Escalation of AI Rivalry
Against this backdrop, geopolitical tensions between the US and China in the AI field are only intensifying. Previously, Chinese authorities declared Anthropic's Claude Code tool a "serious threat" to local users, leading to a ban on its use by Alibaba employees. In response, Washington is launching the Gold Eagle initiative — a platform for sharing vulnerability data among leading US AI companies, including the Department of Defense. The US Treasury Secretary stated an intention to use advanced AI models to outpace adversaries.
The culmination of this policy was the recent decision by the Trump administration to block foreign access to Anthropic's most powerful models — Fable 5 and Mythos 5, which caused a temporary shutdown of these neural networks for all users.
Expert opinion: The situation with Zhipu AI is just the tip of the iceberg. The scale of industrial espionage in the AI field appears to significantly exceed publicly known cases. This tactic of "catch-up development" through outright technology theft not only undermines trust in the Chinese AI industry but also creates serious risks for global security, as models trained on others' data may contain hidden vulnerabilities. The market should prepare for tighter control measures and possibly new sanctions.