Z.ai Claims New Open-Source AI Model Nearly Matches Anthropic in Cybersecurity

Chinese AI company Z.ai has announced its new GLM-5.3 model, claiming that the open-source system can perform close to Anthropic’s restricted Mythos 5 in certain cybersecurity evaluations. The company says GLM-5.3 achieved an 84.5% score on CyberGym, slightly higher than Mythos 5’s reported 83.8%. The comparison is notable because Mythos 5 is designed specifically for advanced cybersecurity work.

However, the results are not uniformly better across all evaluations. On ExploitBench, Z.ai reported that GLM-5.3 scored 54.4%, compared with 78.0% for Mythos 5. This suggests that while the new model may be highly capable at identifying software vulnerabilities, it still has a substantial gap in other cybersecurity-related capabilities. The figures are also company-reported, so independent testing will be important for determining how the models compare in real-world conditions.

Z.ai plans to make GLM-5.3 publicly available in approximately two weeks, but the company is taking a more cautious approach with its most sensitive capabilities. Advanced cybersecurity functions will reportedly be restricted through a trusted-access program, allowing verified users to access them while limiting broader exposure to potentially risky capabilities.

The announcement highlights the growing competition between Chinese and U.S. AI developers. Open-weight models from companies such as Z.ai are increasingly challenging proprietary systems on performance while offering developers greater flexibility. Z.ai’s previous GLM-5.2 model was already positioned as a powerful open model for long-running software-engineering tasks, and GLM-5.3 extends that push into more demanding AI applications.

The release could become another important milestone in the global AI race. If independent evaluations confirm Z.ai’s claims, GLM-5.3 would demonstrate that open-weight models can approach highly specialized closed systems in certain areas. At the same time, the decision to restrict sensitive capabilities shows that greater openness comes with significant safety considerations. As AI models become increasingly capable, developers will have to balance performance and accessibility with responsible controls over powerful applications.