Chinese AI startup Z.ai says its upcoming open-source GLM-5.3 model is approaching the cybersecurity capabilities of Anthropic’s restricted-access Mythos 5, strengthening its position as a lower-cost challenger in advanced AI development.

According to Z.ai, GLM-5.3 scored 84.5% on CyberGym, a benchmark measuring whether models can review code, identify software vulnerabilities and verify that those flaws are real. That slightly exceeded the 83.8% score Z.ai reported for Mythos 5, although the results have not been independently verified.

GLM-5.3 remained significantly weaker in turning discovered vulnerabilities into working exploits. Z.ai said the model scored 54.4% on ExploitBench, compared with 78% for Mythos 5, while Anthropic’s model also completed more attack-development tasks in timed testing.

Z.ai plans to release GLM-5.3 publicly after additional security assessments, while its most sensitive cybersecurity capabilities will be restricted to verified users through a trusted-access program. The company says it has added safeguards to identify risky requests, monitor model behavior and reject malicious tasks.

The approach reflects growing debate over how powerful cybersecurity AI should be distributed. Z.ai argues that advanced defensive tools should remain accessible to open-source developers and smaller security teams rather than being limited to a small number of closed AI providers.

The company is also launching an Open Source Shield initiative that will use its models to audit selected open-source projects and expand security capabilities in its ZCode programming platform.