Chinese AI firms challenge U.S. models in cybersecurity tests

Chinese artificial intelligence firms are advancing capabilities in coding and cybersecurity, challenging American competitors. Zhipu AI recently announced its GLM-5.3 model, which reportedly matched or exceeded the performance of Anthropic’s Mythos 5 model in identifying software security flaws during certain tests. While GLM-5.3 scored higher in coding review capabilities, it was less efficient than Anthropic’s model at converting those flaws into functional software exploits.

Beijing is reportedly adopting a “tiered governance” strategy for open-weight AI models. This approach involves releasing basic capabilities freely while subjecting frontier capabilities to security reviews and restricting highly sensitive models to domestic use. This strategy serves as a hedge against U.S. computing export bans and allows domestic users to run models offline on classified data. Additionally, Moonshot AI recently released its Kimi K3 model with open weights after performing well in developer testing.