A new open Chinese AI model reportedly rivals the performance of Anthropic’s Mythos when it comes to cybersecurity. This further calls into question the utility of US restrictions on that model.
In China, Z.ai has launched a new open AI model: GLM-5.2. This is an open-weight model. In other words, the LLM is freely available for anyone to download and use. Open models typically lag somewhat behind the so-called frontier models from Anthropic and OpenAI in terms of capabilities, but this particular model is said to be among the absolute best regarding cybersecurity.
According to the Wall Street Journal, security researchers have found that it rivals the latest US models, including Mythos, when it comes to identifying security vulnerabilities. In other tasks, GLM-5.2 is said to still perform less effectively.
The ability of an LLM to detect bugs and potentially chain them together to create a vulnerability carries risks. Security teams can use such an AI model to resolve issues in a timely manner, but in the wrong hands, hackers could also abuse it to develop advanced attacks for which no remedy yet exists.
Limitations
The US government has stated that it has therefore imposed export restrictions on Anthropic’s Fable 5 and its Mythos 5 variant. These powerful models are considered too dangerous to release to the general public.
A side note to this is that Anthropic had built in the necessary mechanisms for Fable 5 to prevent abuse, and the government implemented the restrictions unilaterally and without transparent justification. Furthermore, the White House maintains that the LLMs are safe enough for use by people with US nationality.
In any case, the US is using the security argument to slow down the global rollout of AI and keep models within its borders. The fact that a Chinese company is now launching an open model with similar capabilities demonstrates the shaky foundation of that strategy. The US assumes that its own companies are and will remain the absolute leaders in AI, yet Chinese researchers are still managing to catch up. Moreover, they are doing so without access to Nvidia’s most advanced hardware.