Chinese open-weight AI model raises fresh cyber concerns
Zhipu AI’s GLM-5.2 trails top US systems on broad tasks, The Verge reported, but researchers say it can rival Mythos in some bug-finding work.
By Maya Lindqvist · Senior Technology Correspondent
3 min read
China’s Zhipu AI has released GLM-5.2, an open-weight artificial intelligence model that researchers say can perform at the level of Anthropic’s Mythos in some cybersecurity tests, according to The Verge. The claim matters because US officials have been trying to limit China’s access to advanced AI systems and the chips needed to run them.
The Verge reported that GLM-5.2, from Zhipu AI, also known as Z.ai, does not match leading models from Anthropic and OpenAI across broader general-purpose tasks. In narrower work tied to finding software flaws and cybersecurity issues, however, researchers cited by The Verge said the Chinese model has closed much of the performance gap.
The comparison is focused on bug-finding and cyber scenarios, according to The Verge, rather than overall model capability. That distinction is central to the policy concern: systems that can identify vulnerabilities may help defenders secure software, but they may also help attackers find weaknesses faster.
Open weights widen access
GLM-5.2 is an open-weight model, meaning its model weights are available for download and local use, The Verge reported. The International Business Times, cited by The Verge, reported that the model can be run by anyone on readily available hardware.
That design gives skilled users more control than they would have with a tightly managed hosted service, according to The Verge. It also reduces the ability of a company or government to monitor how the model is used once it has been downloaded.
The Verge reported that the same openness that makes GLM-5.2 flexible for power users could also make it easier to misuse. Bad actors could run the model with limited oversight, according to the report.
US officials see national security risk
The Trump administration views Mythos and other advanced AI models that can find vulnerabilities as serious national security threats, The Verge reported. The US government has sought to restrict China’s access to powerful models such as Anthropic’s Mythos and Fable, as well as the hardware needed to train and operate them, according to The Verge.
The release of GLM-5.2 complicates that effort, according to The Verge’s account, because it suggests Chinese developers are narrowing the distance in at least one sensitive area of AI capability. The model’s weaker showing on general tasks does not remove the concern around cybersecurity uses.
OpenAI has also recently introduced GPT-5.6, according to The Verge. The company has limited access to that model after concerns about possible misuse, The Verge reported.
The result is a sharper debate over where AI controls should focus. According to The Verge, US policy attention is increasingly tied to models that can help discover security vulnerabilities, not only to general measures of AI performance.
This story draws on original reporting from The Verge.