Anthropic says it's shutting down millions of distillation accounts per week as Chinese lab accusations escalate

Anthropic shuts down millions of distillation accounts per week as US-China AI tensions escalate

Anthropic's head of national security policy Tarun Chabra accused Chinese AI lab Zhipu AI of distilling Claude and OpenAI models to build GLM 5.2, marking the first time Anthropic has publicly named Zhipu after previously calling out DeepSeek, Alibaba, Moonshot, and MiniMax. The accusation came at the Aspen Security Forum this week.

The scale of the enforcement response is striking. Anthropic is shutting down distillation accounts at a rate of millions per week, according to the segment. The company is also hinting it will expand access to its Mythos models to ensure what it calls a "fair fight for cyber defenders."

Chabra framed distillation as a material national security risk, calling GLM "probably the most advanced Chinese model on the market now" and suggesting it "poses significant cybersecurity challenges." He further implied the US government should do more to restrict Chinese model adoption globally through partnerships similar to efforts against Huawei and ZTE telecom equipment.

The enforcement challenge is real. Distillation can flow through distributed third-party entities—wrapper companies that resell API tokens—making it difficult to track which end customers are actually using the models. Tokens can be passed through multiple systems and jurisdictions, creating a whack-a-mole dynamic that even aggressive account shutdowns struggle to contain.

The timing overlaps with a reported Beijing effort to curb overseas access to Chinese AI models, suggesting a divergence between Chinese government restrictions and company behavior. While Beijing appears to favor tighter controls on domestic model exports, Chinese labs continue to aggressively distill frontier models and release them internationally as open source.