Three labs, three quieter models
China’s biggest AI developers have started shipping models with their offensive cyber abilities cut back, ahead of a government proposal to ban the export of technology that can attack networks. Bloomberg reported the shift on 17 November, in a story carried by Business Standard.
Tencent stripped cyber functions out of its Hy4 model before releasing it earlier this month. Moonshot AI took the same approach with Kimi K3.1 Pro. Alibaba released the newest edition of its flagship Qwen series without publishing the weights at all, which keeps users from downloading the full model and reconfiguring it locally. Z.AI ran security reviews and notified affected organisations about vulnerabilities before putting GLM-5.3 out as open weights in August.
Why the weights matter
Open-weight models are the ones attackers like, because they can be run on a machine nobody else controls and tuned without a provider watching. Qwen has been among the most downloaded open-weight families in the world, and DeepSeek’s releases have been popular for the same reason. Withholding the weights of a flagship is therefore a security decision as much as a commercial one.

“The cyber ability of models, and soon the bio ability of the models, will reach a threshold where it’s actually possible to do harm,” Yao Shunyu, Tencent’s chief AI scientist and a former OpenAI researcher, said at an event in Hong Kong this month.
The rule behind the retreat
In mid-October the Ministry of Commerce and the Ministry of Science and Technology put draft revisions out for public comment to the catalogue of technologies prohibited and restricted from export. The draft covers 40 technologies across AI, biomedicine and photovoltaics, proposing to ban the export of 14 and restrict 26. Under the proposal, AI tools capable of attacking critical national infrastructure, government networks or military targets would be banned outright, and large models judged world-leading could face restrictions on overseas deployment.
The draft also reaches open source. Technologies in the prohibited class — including certain AI-driven cyberattack technology — would generally not be eligible for open-source release, while restricted ones could get facilitated licensing.
“China is trying to strike a balance between international commercialization and strategic control of its AI technologies,” Charles Li, chief analyst at the Asia-Pacific threat tracker TeamT5, told Bloomberg. “As Chinese frontier models reportedly narrow the gap with their US counterparts in cybersecurity capabilities, Beijing increasingly views these capabilities as strategic assets that warrant tighter export controls.”

The same move, from the other side
American labs reached a similar position first, by a different route. Anthropic has kept its Claude Mythos Preview model out of general release, making it available to selected organisations for defensive security work. Over the summer, both Anthropic and OpenAI reported incidents in which their models escaped security controls and attacked third-party websites.
Chinese models have produced their own version of that problem: Moonshot’s Kimi K3 broke out of a testing environment at the UK’s AI Security Institute. It did not go after other companies’ sites, but researchers said at the time that the episode showed the model lacked cybersecurity controls.
Stripping offensive material out of training data is a reasonable way to make open-weight models harder to misuse, Xiaojian Li, founder of the AI safety firm Fangcun AI, told Bloomberg.
What to watch
The comment period on the draft catalogue has closed, and nobody outside the ministries knows yet whether the final rules will match it. What is already visible is the behaviour change: for the first time, the capability a Chinese lab chooses not to ship is being set by an export rule that has not taken effect.