CHINESE AI
Washington has a new AI headache
Chinese AI models are catching up with top US systems in some cybersecurity tasks, adding fresh pressure to the global AI race.
Researchers say Zhipu AI’s new GLM-5.2 model can match leading US models at finding software bugs, although it still falls behind Anthropic and OpenAI in other areas.
The model is open-weight, meaning people can download it, run it on their own hardware and modify it with fewer restrictions.
That makes it useful for companies that want cheaper, more flexible AI.
It also makes security experts nervous, because hackers could use the same tools to spot software weaknesses before they are fixed. Delightful little nightmare.
GLM-5.2 has already become one of the most-used AI models on OpenRouter.
Cybersecurity firm Semgrep said it beat Anthropic’s Claude Opus 4.8 in some tests, while researchers said both GLM-5.2 and Anthropic’s models can match Mythos in certain bug-finding tasks.
In brief:
Chinese AI models are narrowing the gap with US systems in some cybersecurity tasks.
Open-weight models are cheaper and more flexible, but harder to control.
US restrictions could push more companies towards Chinese alternatives.
Cyber stakes rise
The timing is awkward for the US. Washington has been tightening access to some advanced AI models over security concerns, including recent limits on OpenAI and Anthropic systems.
Critics say this could push more companies towards cheaper Chinese open-weight models instead.
The wider issue is no longer just who has the best model. It is who gets access, who controls it, and whether open-weight AI will shift power in cybersecurity.
Twas ever thus. - MV