China's Z.ai's new model matches Mythos 5 in cyber-defence tests

China's Z.ai said its new open-source GLM-5.3 model neared Anthropic's restricted Mythos 5 in identifying software vulnerabilities. It claimed GLM-5.3 scored 84.5% on CyberGym, which tests a model's cybersecurity capabilities, compared to 83.8% reported for Mythos 5. The model trailed Mythos 5 in converting discovered flaws into working attacks, part of defensive security research.