Experts already claim that the Chinese GLM-5.2 model is as “dangerous” as Anthropic’s

The technological gap between the US and China continues to narrow. At least, if we pay attention to what they say the latest analyzes on the GLM-5.2 model. Two independent cybersecurity companies have made their own assessment and their data reveals that in terms of cybersecurity, GLM-5.2 is as good as Claude Opus 4.8. That has notable implications, especially considering how the US government is now restricting access to Anthropic and OpenAI’s frontier models.

AI in the face of the threat of cybersecurity. Since Claude Mythos Preview appeared, the discourse on AI has changed significantly. Suddenly the world realized that these models could become weapons with which to find vulnerabilities in all types of systems to exploit them. Anthropic has already warned that Mythos was too dangerous to be publicly available, and it did not matter that it released hidden versions like Fable 5 shortly after: the US Government has temporarily vetoed them and the same has happened with GPT-5.6. The situation is unusual.

Beware of Tulongfeng (or not). Last Wednesday, a Chinese cybersecurity company called 360 Security Technology (Qihoo 360) launched a new vulnerability detection tool called Tulongfeng. According to its creatorsTulongfeng is comparable Mythos in this task. The company this on the US “Entity List” since May 2020 and its CEO, Zhou Hongyi, stated that Mythos is equivalent to a “cybernuclear weapon.”

The Sputnik moment with GLM-5.2. But the real recent protagonist of the Chinese AI industry is the GLM-5.2 model from the startup Zhipu.ai (Z.ai), which is becoming very popular by demonstrating performance comparable to the best models from US companies. Its fundamental advantage is that it is an open weights model: any person or company can download it, modify it and run it on their own hardware (although it requires a huge amount of video/unified memory to be able to use it, it is a model with 744B of parameters). But he is not only good at programming or at agentic tasks.

Better than Claude in cybersecurity? The cybersecurity firm Semgrep stated in a recent analysis that GLM-5.2 was superior to Claude Opus 4.8 regarding cybersecurity and pointed out that “we have a Mythos at home.” Another independent study from Graphistry stated basically the same thing when comparing it with Opus 4.8 and GPT-5.5. Not only that, it achieved excellent results at a fraction of the price: one-sixth of what it cost to run tests with Claude Opus 4.8, for example. Axios revealed little cited a cybersecurity researcher who explained that GLM-5.2 is capable of chaining exploits “in the same way that an elite human attacker would.”

Chinese mythos before 2027. Jie Tang, CEO of Z.ai, responded to a Twitter thread in which it was pointed out that at this rate, China would have an AI model at the level of Mythos or Fable by the end of 2026. Elon Musk himself intervened saying that in his opinion this Chinese model with such performance would arrive in the first quarter of 2027. Jie Tang was forceful and replied to Musk saying “it won’t take that long.”

And we also have Sakana Fugu. These days we also learned the news that Sakana AI, a Japanese AI startup, had launched Fuguan AI model that is actually not so much an AI model as it is a router or orchestrator of other models. What it promises It is to perform at the level of the best models in the US, taking advantage of different models, both open and closed. The internal benchmarks are promising, but some independent analysts they explained Although the idea is not bad, its performance and cost are not as striking as the company claims.

While the US blocks its models, China advances. The situation is paradoxical, because what China is doing is precisely taking advantage of a unique moment. The US is restricting the deployment of the most advanced AI models from Anthropic and OpenAI to avoid cybersecurity risks. And while that happens, Chinese companies are apparently closing the gap with truly remarkable open models.

In Xataka | The prompt engineering fashion is over. Now what is important is loop engineering