Moonshot AI, Tencent and Alibaba, three of China’s leading model developers, have each done the same thing with a new model in the past month: pushed offensive cyber capability down while keeping, or even strengthening, defensive capability. Bloomberg strung the releases together this week and linked them to the draft AI export-control rules the Ministry of Commerce published in October. None of the three companies has publicly tied its approach to the draft.

What each company did
Moonshot AI moved first. When it released Kimi K3.1 Pro in late October, it said it had removed cybersecurity-related data from the training corpus and the reinforcement-learning environments, while leaving general coding ability intact. On the company’s internal offense-and-defense benchmark, which tests finding and exploiting software vulnerabilities in a controlled environment, K3.1 Pro scored 81.0%. A comparable version without the risk reduction scored 90.5%, and the 81.0% still sits somewhat above the previous generation. These figures have not been independently verified.
“We aim not to release open weights that would extend the absolute frontier of cyber capability.” — Moonshot AI
Tencent released Hunyuan Hy4 on November 4. It says the model’s cybersecurity training was used only for defense, such as finding software flaws and helping fix them, and that material on building hacking tools, breaking into systems and probing system weaknesses was excluded. After testing by in-house security experts, Tencent says it saw no elevated risk of Hy4 helping people break into software or bypass security protections, and found no sign that it makes harmful biological experiments easier. Tencent plans to release the weights later, together with a safety testing report.
Alibaba released Qwen4-Max on November 16. It says safety checks were added during training to reduce vulnerabilities such as code injection, memory-safety bugs and broken access control, and that protections against misuse for cybercrime and dangerous substances were also strengthened. Internal testing found no clearly high-risk capabilities, it says. Bloomberg’s report adds that this flagship generation ships without open weights, so users cannot download and modify a full version themselves; Alibaba did not comment separately on that point.
Where the draft draws its lines
The draft, published jointly by the Ministry of Commerce and the Ministry of Science and Technology in October, would put 23 AI technologies on an export-control list, covering large models, agents, chip design and compute clusters. For large models, the controls apply to core components such as weights and source code, with a threshold of at least 2 trillion parameters, or a top-ten global ranking in performance, safety and similar tests. The risks the draft names include AI-driven generation and mutation of malware and AI-assisted vulnerability-discovery tools that could be used to attack critical infrastructure, the defense industry and core government systems.
The exemptions are spelled out too. Open-source code and published research fall outside the controls. Serving overseas users through apps, websites and APIs is still allowed. Running a controlled model on overseas servers counts as an export, but companies running it on their own overseas cloud servers automatically receive a license. Public comment closed on November 15; no date has been announced for the final version.
Measured against the threshold, Kimi K3.1 Pro describes itself as being in the 3-trillion-parameter class, and Qwen4-Max has a backbone of 2.4 trillion parameters. Both sit above the 2-trillion line.
What it means for developers and security teams
If the final rules are close to the draft, two groups would feel it most directly.
The first is teams building on open-weight models. For models above the threshold, whether weights can keep going out to overseas users depends on licensing, and whether they can still be downloaded freely inside China is not addressed directly in the draft; that will have to wait for the final text and follow-up guidance. Moonshot AI has already chosen to cut capability first and release weights afterward, and Tencent says its weights will come later. The next round of releases should show whether those two paths become the industry default.
The second is security practitioners who do penetration testing and red-team exercises. Using open models to help write exploit code or run vulnerability checks has been common, and newer Chinese models will be more conservative here. For defensive work such as reviewing code, finding flaws and writing patches, all three say capability has been kept or improved; real-world results will have to wait for third-party evaluations.
There is one point of comparison. In the United States, OpenAI’s approach is to make its more capable cybersecurity models available only to vetted defenders through a screened-access program. The path Chinese developers have publicly described is to strip out offensive capability at the training stage. For now, the effectiveness of both approaches rests on the companies’ own statements.
Sources: Bloomberg, Reuters coverage of Tencent Hy4 and Kimi K3.1 Pro, CocoLoop, The Next Web. Benchmark scores follow Moonshot AI’s own testing; the number of controlled items and the parameter threshold follow the draft for comment from the Ministry of Commerce and the Ministry of Science and Technology.