China's Z.ai Claims AI Model Nears Anthropic's in Cyber-Defense Tests

Chinese AI startup Z.ai is positioning its new open-source GLM-5.3 model as a significant challenger to Western AI security tools, claiming it nearly matches Anthropic's restricted Mythos 5 in cybersecurity tests. The company announced Friday that GLM-5.3 achieved an 84.5% score on CyberGym, a benchmark for identifying software vulnerabilities, marginally surpassing Mythos 5's reported 83.8%.

However, these results have not been independently verified. While GLM-5.3 showed strength in identifying flaws, it lagged behind Mythos 5 in the crucial step of converting discovered vulnerabilities into working exploits. On the ExploitBench test, GLM-5.3 scored 54.4%, compared to Mythos 5's 78.0%. In timed tests, Mythos 5 also outperformed GLM-5.3 in completing attack-development tasks.

Anthropic has limited access to Mythos, a cybersecurity-hardened version of its Claude Fable 5 model, to vetted organizations. This approach stems from concerns that AI capable of finding and exploiting software flaws, while useful for defenders, could also lower the barrier for malicious actors.

“

To the best of my knowledge, this is the first time a Chinese lab is publicly justifying a delayed open release of model weights with safety considerations.

Z.ai intends to release GLM-5.3 publicly in approximately two weeks, following security assessments and the strengthening of its safeguards. The company stated that its most sensitive cybersecurity functions will be accessible only through a "trusted access" program for verified users. This strategy echoes Anthropic's "Project Glasswing" for Mythos, suggesting a growing trend toward responsible release practices even for open-weight models.

Gabriel Wagner, an AI governance researcher at Concordia AI, noted this as potentially the first time a Chinese lab has publicly cited safety considerations for delaying an open-weight model release. He suggested this indicates increasing sophistication in risk management practices within China's AI sector. Z.ai emphasized its implementation of multiple protection layers, including request screening and monitoring, to differentiate harmful activities from legitimate cybersecurity tasks.

“

This shows that open-weight risk management practices in China are becoming more sophisticated.