China's latest open-weight model rivals U.S. models at hacking - axios.com

Axios reported that China-based AI lab Z.ai has delayed the public release of the weights for its latest open-weight model, GLM-5.3, by two weeks. The lab said the delay is necessary because the model is so capable at finding and exploiting security flaws. Z.ai's warning indicates the company intends to use the additional time to test safety and security controls before the weights are distributed.

The article quotes Z.ai's warning that GLM-5.3 is "so capable" at identifying and exploiting security flaws that the two-week postponement was required. During this period, the lab said, it will test safety and security controls. The article also states that Chinese open-weight models are "closing in" on U.S. frontier models' ability to find and exploit security flaws. The title of the piece describes the model as rivalling U.S. models at hacking. The article does not specify which U.S. models are referenced.

The model is described as open-weight, meaning its trained parameters are intended to be released publicly. Open-weight releases allow broad access to the model, including by researchers and other third parties. The article's framing suggests that the capability gap between Chinese open-weight models and U.S. frontier models in security flaw identification is narrowing. The two-week delay is presented as a precaution by Z.ai in advance of a public release. The article does not state whether other Chinese labs have taken similar steps.