Bloomberg reports that leading domestic AI companies such as Alibaba, Tencent, and Dark Side of the Moon have intentionally weakened offensive network capabilities in their new generation models while improving defensive performance.
A specific change: Alibaba’s latest Qwen flagship does not have open weights, and users cannot download the full version and modify it themselves.
Before Tencent released Hunyuan Hy4 this month, it also cut off a round of network-related functions.
The background is the export control draft released by the Commerce Department last month, which intends to ban the export of AI tools that can attack critical infrastructure and government networks.
It’s even more interesting to look at this week: California in the United States just sued OpenAI for losing control of its agent, Gemini 4 Argon was exposed to hit real websites by itself during testing, and the FBI is notifying the open source attack framework Artex. Model manufacturers in China and the United States began to tighten the same type of capabilities at almost the same time.
One thing needs to be made clear: it is the offensive capabilities that are weakened, not the safety capabilities. When it comes to defensive tasks such as finding vulnerabilities, writing patches, and analyzing logs, several companies say their performance is improving. The difficulty is that the two are often the same ability in the model. How to draw the boundary between offense and defense has not been disclosed by each company.
The impact on developers is mainly on the open source side. The lack of weight on the flagship means that teams that want to use the strongest domestic models for localized security research can only use the API or fall back to the previous generation open source version.