Z.ai GLM-5.3 Hits 84.5% on CyberGym, First Chinese Lab Delay for Emergent Safety

Post-training optimizations enabled the 743B parameter model to outperform proprietary systems including GPT-5.6 Sol and Mythos 5 in vulnerability detection. The model's coding performance rose 50% over its predecessor, GLM-5.2, while its ExploitBench score climbed from 24.4% to 54.4%. Z.ai discovered the system had identified 2,436 vulnerabilities in 269 different projects, some dating back 40 years.

Z.ai withheld the open weights for a 2 week period to conduct safety hardening and vetting by security partners after the model developed a high capacity for targeting potential exploits. The firm stated that the cyber reasoning was not explicitly programmed but emerged during fine-tuning. This is the first time a Chinese laboratory has cited emergent AI capabilities as the specific reason for a release delay.