Chinese Startup Z.ai Claims New AI Model Rivals Anthropic's Mythos 5 in Cyber-Defence

Chinese AI startup Z.ai announced that its open-source GLM-5.3 model has nearly matched the performance of Anthropic's restricted Mythos 5 in identifying software vulnerabilities. According to Z.ai, GLM-5.3 scored 84.5 percent on the CyberGym test, which evaluates a model's ability to review code and identify security flaws, compared to 83.8 percent for Mythos 5. These results have not been independently verified.

Despite the strength in identification, Z.ai reported that GLM-5.3 lagged behind Mythos 5 in converting those flaws into working attacks. On the ExploitBench test, GLM-5.3 scored 54.4 percent while Mythos 5 scored 78.0 percent. In timed attack-development tasks, Mythos 5 completed significantly more tasks than GLM-5.3 over both two hour and six hour periods.

Z.ai plans to release GLM-5.3 publicly in approximately two weeks following security assessments. The company stated that its most sensitive cybersecurity functions will be limited to verified users through a trusted access programme. This approach mirrors the limited-access scheme used by Anthropic for Mythos, which is a version of the Claude Fable 5 model with cybersecurity safeguards removed.

Z.ai framed the launch as a challenge to closed-source models, arguing that advanced cyber-defence tools should be available to smaller security teams and open-source developers. The company also announced an Open Source Shield initiative to audit selected open-source projects. Gabriel Wagner, a researcher at Concordia AI, noted that this marks the first time a Chinese lab has publicly justified a delayed open release based on safety considerations, suggesting that risk management practices in China are becoming more sophisticated.