Anthropic reported on September 29 that Z.ai's GLM-5.3 approached its Mythos model at certain cyber exploitation tasks. Since August 28, GLM-5.3's weights have been freely downloadable on HuggingFace, with safeguards that can be bypassed or removed. The article argues that China's "development" and "security" balance, in place since Li Qiang's 2024 "AI Plus" initiative, will likely push Chinese labs to restrict access to their most disruptive models while continuing open-source diffusion at the low end.
No score is assigned. Sources and their independence are shown in the citation chain below.