Chinese AI lab Z.ai gets warning similar to top US AI companies

China's AI research lab Z.ai has issued a warning similar to that of leading US AI companies like OpenAI and Anthropic. The lab's latest model, GLM-5.3, is so proficient at identifying and exploiting security vulnerabilities that Z.ai plans to postpone the public release of the model weights for two weeks to bolster safety measures.

The company is currently testing and enhancing security controls before wider availability. This move follows similar concerns raised by Sam Altman of OpenAI and Dario Amodei of Anthropic, who have expressed apprehensions about the potential misuse of powerful open-weight models, particularly those capable of cyber exploitation.

Z.ai's GLM-5.3 has demonstrated significant capabilities in vulnerability discovery, outperforming models such as Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol on certain benchmarks. The lab has trained GLM-5.3 to excel in detecting security flaws, with the model already uncovering over 2,400 vulnerabilities, including critical and high severity ones, in systems like the Linux kernel and popular software like VMware and Apache.

Z.ai views its release of GLM-5.3 as primarily beneficial for defenders rather than a security risk, emphasizing the need for an open shield alongside open attack surfaces. The company has also introduced a program allowing open-source maintainers to use GLM-5.3 to scan their repositories for bugs. However, Z.ai acknowledges that once the model's weights become public, the company will have limited control over how others modify or utilize the model, a common challenge with open-weight releases.

The US Trump administration is reportedly considering regulatory approaches to address the growing capabilities of open-source AI models, which are increasingly comparable to proprietary counterparts.