China’s Z.ai unveils new AI model nearing Anthropic’s Mythos 5 in cyber tests

On Friday, Chinese AI startup Z.ai announced that its open-source GLM-5.3 model had approached Anthropic’s restricted Mythos 5 in detecting software vulnerabilities, strengthening the case for a Chinese AI challenger securing a foothold with Western developers.

According to Z.ai, GLM-5.3 scored 84.5% on CyberGym in a crucial test of whether a model can review and scrutinize flaws that are real.

The results were marginally higher than the 83.8% reported for Mythos 5. However, GLM-5.3 fell behind Mythos 5 in transforming exploited vulnerabilities into working attacks as part of a standard defensive security research. Z.ai underlined that its model scored 54.4% on the ExploitBench test for this capability.

Anthropic has developed a version of its Claude Fable 5 model, Mythos with cybersecurity protections removed, and only available to certified organizations.

These controls reflect concerns that AI systems capable of finding and exploiting software faults could assist defenders but might also enable threats.

Regarding the official release, Z.ai announced it would release GLM-5.3 in about two weeks after complete security assessments.

Its most critical security capabilities and functions would be available only to verified users through a trusted access programme. Z.ai says cybersecurity capabilities were expanded post-training as a general programming model rather than developed as a tailored security system.

Nonetheless, the launch built on the international interest in GLM-5.2 gained traction in recent times for coding and agent capabilities that users claim approach those of leading US models but at a lower cost.