The South China Morning Post reported that leading Chinese AI companies are starting to limit the hacking skills of their strongest open-weight models. Open-weight means anyone with enough hardware can download and change the model. The firms are removing hacking-related material from the data used to train the models. The paper ties the move to several high-profile AI hacking incidents involving US and Chinese firms. In September, Anthropic reported that Z.ai's GLM-5.3 could build sophisticated exploits on its own, nearing Anthropic's restricted Claude Mythos Preview.
Tencent and Moonshot AI both released their latest models in stages, with the open version arriving weeks after the closed one. Tencent has not yet released the open version of Hy4 and says it will come soon. Moonshot said its filtered model scored 81 per cent on an internal cyber test, against 90.5 per cent for an unrestricted version. Experts quoted say the method is not a full fix, because skills that help defenders can also help attackers. Analysts call it an emerging practice, not a standard. One analyst said there is no evidence yet of a wider loss in competitiveness. Whether it lasts is still open.
Anyone choosing open-weight models for their own systems is relying on limits that experts call helpful but incomplete. A skill a model never learned is harder to restore than one it was only trained to refuse, one security executive said, but not impossible.