A new Chinese AI model, GLM-5.2, is attracting attention because experts say it performs almost as well as some of the world's best AI systems on cybersecurity tasks.

According to researchers at Graphistry and Semgrep, the model can help identify software issues and support cybersecurity investigations at a much more affordable price than main players like OpenAI, Google, or Anthropic.

The model was released by Chinese startup z.AI (formerly Zhipu Ai) and has quickly caught the attention of the security community. Developers say it is on par with US models for carrying out tasks like vulnerability discovery and security investigations, and at around half the cost.

What makes it different is that people can download it and run it on their own computers.

As more commercial models have built-in guardrails that could suspend a user for abuse, GLM-5.2 can be downloaded onto a user's computer, allowing users to operate with free range, and potentially remove safety measures.

If used legitimately for cybersecurity research, it could mean a more affordable option for the industry, but without diligence and industry oversight, GLM-5.2 could get into the wrong hands.

That means criminals have access to extremely powerful AI tools behind closed doors – capabilities once limited to a handful of companies are becoming easier for almost anyone to access.

In the eyes of some, this has been coming for a while. "China's release shouldn't come across as a shocking breakthrough," says Gene Moody, field CTO at cybersecurity firm Action1.

“Once a capability has been demonstrated, others will inevitably replicate it," Moody said.

Why security experts are concerned

AI can already help criminals write convincing phishing emails, search for software flaws, and speed up parts of a cyberattack. A model that anyone can download gives cybercriminals more freedom to experiment without being monitored.

Axios reports that hackers are already discussing ways to remove the model's safety restrictions, and some researchers found that simple prompts were enough to persuade the AI to answer questions it normally shouldn't.

Unlike cloud-based AI services, there's no company watching how the model is being used once it's running on someone's own computer.

That makes it much harder to stop bad actors from adapting the technology for malicious purposes. And as Moody explained, restricting access can backfire by encouraging developers to develop their own versions instead.

An AI arms race

AI isn't replacing hackers yet, and today's AI-generated attacks still have plenty of limitations. But the technology is improving quickly and becoming more affordable every year.

The bigger concern is that countries and companies are now racing to build ever more powerful cyber AI. That means each breakthrough is likely to be copied by competitors.

"The concern isn't really that China has produced a competitive cybersecurity model," Moody says.

The Cybernews community is talking about this. Be a part of the conversation.

"AI is rapidly becoming another arena for geopolitical competition instead of collaborative defense."

Moody warns that every new model pushes the race forward rather than solving the underlying problem.

"Each new model will simply represent another turn in an arms race where no one truly wins."

For everyday users, that could mean defenders and cybercriminals both gain access to increasingly powerful AI tools, and the race between them intensifies.

---