China's AI labs are quietly teaching their models to forget how to hack
Alibaba, Tencent, Moonshot and Z.AI have all dialed back offensive cyber skills, Bloomberg reports, a month after Beijing drafted rules that treat hacking ability as a strategic asset.
For the past two years, China's AI pitch to the world has been generosity. Release the weights, let anyone download the model, run it on their own hardware, undercut the Americans on price. It won a lot of developers. It also meant that anything those models could do, anyone with a GPU could make them do.
That includes hacking, and Beijing has noticed.
Bloomberg reported this week that China's top AI developers are now releasing models with weaker offensive cyber capabilities and better defensive ones. The details vary by company. Tencent cut cyber functions from its Hy4 model before releasing it earlier this month. Moonshot AI took a similar approach with Kimi K3.1 Pro. Z.AI ran security reviews and warned organizations about vulnerabilities before it made GLM-5.3 open-weight in August. And Alibaba, which built much of its reputation on open releases, put out the latest edition of its flagship Qwen series without downloadable weights at all.
The method most of them seem to favor is simple: strip offensive security material out of the training data so the model never learns it properly. Xiaojian Li, founder of Fangcun AI, told Bloomberg it was a sensible way to make open models harder to abuse, though he didn't oversell it. "There is no such method now that can completely eliminate models' risk," he said.
Why now
The timing isn't subtle. On October 16, China's Ministry of Commerce and Ministry of Science and Technology published a draft revision to the country's technology export control list. It covers 40 technologies, 23 of them in AI, spanning large models, AI agents, computing clusters and chip design. According to TNGlobal, examples on the prohibited side include systems that autonomously generate or modify malicious code. AI-assisted vulnerability discovery would be restricted. The rest of the list covers five biomedicine technologies and 12 in solar manufacturing, and across all 40 items the draft would ban 14 from export and restrict the other 26.
The draft goes further than hacking tools. Bloomberg reported in October that it would restrict exports of core technology, meaning weights, checkpoints, source code and architecture, for models with at least 2 trillion parameters or models ranked in the world's top 10 on performance, usage and safety. Bloomberg gave Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max as examples of models big enough to cross the 2 trillion line. Models trained on "core data" under China's Data Security Law would be banned from export outright. A commerce ministry spokesperson said the government's management of these technologies "aims to solidify the security baseline."
There are big carve-outs. Technology already public through open-source releases or academic papers isn't covered, and serving a model overseas through an API, website or app generally doesn't count as exporting it. Public comments closed on November 15. A temporary registration requirement for covered exports has been in force since the day the draft appeared, and Chinese firms that run their own overseas cloud or compute centers may get simplified or automatic licensing. In early November the National Development and Reform Commission added that it was speeding up new AI safety standards and ethics guidelines.
Put the draft next to the model releases and you can see how Beijing now thinks about its own labs. Charles Li, chief analyst at the Taiwanese security firm TeamT5, put it plainly to Bloomberg: "Beijing increasingly views these capabilities as strategic assets that warrant tighter export controls."
The DeepSeek problem
There's a recent reason for the anxiety. Bloomberg notes that CrowdStrike said last month that hackers used DeepSeek models in attacks on organizations in 14 countries. Open weights are the hardest case for any safety team, because once a model is on someone else's machine, its maker loses every control it had. No usage policy, no monitoring, no off switch.
Bloomberg also reports that Moonshot's Kimi K3 escaped a test environment at the UK AI Security Institute, which makes it hard to argue that Chinese models are too weak to worry about.
Tencent's chief AI scientist, Yao Shunyu, who used to work at OpenAI, said at an event in Hong Kong this month that the models "will reach a threshold where it's actually possible to do harm." His employer cut the cyber functions out of Hy4 before it shipped, so you can see how seriously the company is taking that threshold.
American labs went through the same reckoning first. Anthropic has kept its Mythos Preview model out of general release and offers it only to selected organizations for defensive security work. OpenAI and Anthropic both reported incidents this year in which their models slipped security controls during testing. The difference is that in the US those decisions were mostly made by companies, under pressure, one at a time. In China the state wrote the rule first and the labs fell into line within a month.
Whether it works is another question. Stripping training data can blunt a model, but it doesn't stop someone fine-tuning the skill back in, and the older open models that already know how to hack aren't going anywhere. China has decided its most dangerous models are worth keeping at home. It still has to deal with the ones it already gave away.
Sources: Bloomberg via The Star, TNGlobal, Taipei Times (Bloomberg).
More from the feed
A judge compared police face matching to a Magic 8 Ball. Fargo wants immunity anyway
A federal judge in Kentucky says the Fourth Amendment won't let a detective hand probable cause to AI. A week later, Fargo answered a Tennessee grandmother who spent five months in jail on a face match, and asked to have her case thrown out.
Washington closed the cloud loophole on AI chips. It also stopped calling it AI
A new Commerce Department rule means a data center anywhere on Earth needs a US license if it rents serious compute to Chinese companies. The 132-page text also swaps "artificial intelligence" for "Super Intelligence," by executive order.
Hong Kong switched on face recognition today. Police won't say what it can do
From Monday the city's SmartView cameras can pick a face out of the footage and follow it. The police chief says the tool won't be abused. He also says its capabilities are a secret.