Chinese labs that sell open-weight models abroad are cutting offensive cyber skills before MOFCOM's catalog is final, making cyber performance a release decision with export consequences.
China’s leading model developers are shipping their newest systems with weaker offensive hacking skills, as the Ministry of Commerce (MOFCOM) prepares to bar exports of AI tools built to attack critical infrastructure. Bloomberg reported on November 16 that recent releases from Alibaba, Tencent and Moonshot AI have scaled back offensive cyber capabilities while strengthening defensive ones.
The individual steps were partly visible before. What Bloomberg adds is the pattern: three of China’s most closely watched labs moving the same way within weeks of a draft export rule that names AI-driven malware as a banned technology. Alibaba released the latest edition of its Qwen flagship without its weights, the internal settings that let users download and customize a full model. Tencent cut cyber functions from Hy4 before its release earlier this month, and Moonshot took a similar line with Kimi K3.1 Pro.
None of the companies has published an independent audit of how much offensive capability was removed. Their own figures and statements are, for now, the only public measure.
MOFCOM’s draft puts AI attack tools on the banned list
On October 16, MOFCOM and the Ministry of Science and Technology opened consultation on revising the Catalog of Technologies Prohibited or Restricted from Export (《中国禁止出口限制出口技术目录》). The ministry’s spokesperson Q&A said 23 key technologies related to large models, AI agents, computing clusters and information processing would be added, within a wider package of 14 proposed prohibited and 26 proposed restricted technologies.
AI-driven malicious code generation and mutation technology (人工智能驱动的恶意代码生成与变异技术) sits on the prohibited side. MOFCOM said it could be used to attack critical information infrastructure, defense-industry systems and core government systems. AI-assisted vulnerability discovery is proposed for the restricted list, and the ministry said deploying restricted AI technology locally overseas would count as a technology export. Bloomberg added that "world-leading" large language models could face limits on overseas deployment under the draft.
The joint announcement took public comments from October 16 to November 15, so the comment window closed a day before Bloomberg’s report. It also put temporary registration in place: until the revised catalog takes effect, exporters of listed technologies must register through MOFCOM’s technology trade system before export. EastFrontier set out the commercial trade-offs of the draft when it was published.
Labs are removing attack material from training, at a measurable cost
Moonshot gave the clearest numbers. According to a Reuters report on October 24, the company stripped cyber-related data from both the training corpus and the reinforcement-learning environments for Kimi K3.1 Pro while keeping general coding skills. Moonshot’s own exploit test put Kimi K3.1 Pro at 81.0%, against 90.5% for a comparable build trained without the cuts. Those results have not been independently verified. "We aim not to release open weights that would extend the absolute frontier of cyber capability," Moonshot said.
Tencent narrowed Hy4’s cybersecurity training to defensive work such as finding and fixing software flaws, and left out material on building hacking tools, breaking into systems and probing them for weaknesses, Reuters reported on November 4. The company said "we are committed to a more responsible approach to model training, open-sourcing, and release," and plans to publish Hy4’s weights later with a safety-testing report.
Zhipu, which trades as Z.ai, took a different route with GLM-5.3. In August it said the most sensitive cybersecurity functions would be open only to verified users through a trusted-access program, according to Reuters’ August report. Z.ai also claimed a CyberGym score of 84.5%, slightly above the 83.8% it cited for Anthropic’s Mythos 5. Bloomberg said Z.ai ran security reviews and alerted organizations to vulnerabilities before making GLM-5.3 open-weight.
Open weights are what make these choices consequential. Hackers can run a downloaded model locally and change its settings, Bloomberg noted, and CrowdStrike researchers said last month that attackers used a DeepSeek model against organizations in 14 countries.
Beijing and the labs treat cyber skills as an export asset
Tencent’s chief AI scientist, Yao Shunyu, said at a Hong Kong event this month that "the cyber ability of models, and soon the bio ability of the models, will reach a threshold" at which harm becomes possible, according to Bloomberg. Charles Li, chief analyst at TeamT5, told Bloomberg that "China is trying to strike a balance between international commercialisation and strategic control of its AI technologies."
Filtering has limits. Xiaojian Li, founder of Fangcun AI, called the approach "a good balance between cost and effectiveness, but it’s not perfect," adding that "there is no such method now that can completely eliminate models’ risk." Bloomberg also recalled that Moonshot’s earlier Kimi K3 broke out of a testing environment run by the UK government’s AI Security Institute.
The National Development and Reform Commission said in early November that China would speed up AI safety standards and ethics guidelines, Bloomberg reported. With comments closed, the next marker is MOFCOM’s final catalog: whether its thresholds for "world-leading" models and vulnerability-discovery tools match the cuts the labs have already made, or force deeper changes to open-weight releases such as Hy4’s promised weights.