Z.ai's GLM-5.3: Near Mythos 5 in Cyber Security Tests

Chinese AI startup Z.ai says its new open-source model, GLM-5.3, performs similarly to Anthropic’s Mythos 5 in some cybersecurity tests. Z.ai reported that GLM-5.3 scored 84.5% on CyberGym, an industry benchmark for vulnerability detection. Anthropic’s Mythos 5 scored 83.8% on the same test. However, GLM-5.3 lagged in active exploit generation. It achieved 54.4% on the ExploitBench test, while Mythos 5 scored 78%. Also, GLM-5.3 completed 130 attack-development tasks in six hours, compared to 247 for Mythos 5. Z.ai plans a public release of GLM-5.3 in about two weeks after safety audits. But its most sensitive cybersecurity features will be restricted to vetted partners. This phased rollout is seen by some as a step forward for risk management in China. This news highlights the ongoing competition and evolving safety measures in the AI cybersecurity space.