Reports describe different controls aimed at limiting offensive cyber capabilities in Chinese AI models. Alibaba reportedly withheld Qwen4-Max’s public weights, Tencent reportedly excluded offensive material from Hy4’s training, and Moonshot AI reportedly removed cyber-related data from Kimi K3.1 Pro’s training. Moonshot also reported a numerical comparison from an internal test.

What did each company reportedly change?

Company and modelReported control
Alibaba, Qwen4-MaxAlibaba reportedly withheld the model’s public weights. It also described filtering training tasks by purpose, retaining defensive work and removing offensive tasks.
Tencent, Hy4Tencent reportedly excluded offensive exploitation material from training while retaining defensive security work.
Moonshot AI, Kimi K3.1 ProMoonshot AI reportedly removed cyber-related data from training and reinforcement learning while retaining general coding ability.

Withholding weights affects access: users could not download and customize the complete Qwen4-Max model. Training filters operate at a different stage, by excluding selected material as the model is trained. The reported measures therefore target different parts of model distribution and development.

What do Moonshot’s test scores show?

Moonshot reported an internal exploit-test score of 81.0% for filtered Kimi K3.1 Pro, compared with 90.5% for a comparable build without the data cuts. That is a difference of 9.5 percentage points. The figures are Moonshot’s internal results, not an independent evaluation.

NeoTeo previously covered Moonshot AI’s review of Kimi security findings; that article concerns a different development from the reported training changes to Kimi K3.1 Pro.

China’s export-control proposal is separate context

China’s Ministry of Commerce published a draft export-control proposal on October 16, 2026, that included AI-related technologies. The reported model changes concern decisions by Alibaba, Tencent and Moonshot AI.

A broader survey by The Frontier Wire found other Chinese developers continuing to train or promote cybersecurity capabilities.