Z.ai's GLM-5.3: 84.5% on CyberGym, Release Delayed

Z.ai's GLM-5.3 model has achieved an 84.5% score on the CyberGym vulnerability benchmark. This performance surpasses its predecessor, GLM-5.2, and other proprietary models. Engineers enhanced the model by fine-tuning its agentic capabilities and optimization routines, without altering the base weights. This improvement was so effective at finding and exploiting vulnerabilities that Z.ai has delayed its open release. The company is conducting further safety reviews of the model's agentic capabilities. This development highlights the ongoing challenge of balancing advanced AI performance with necessary safety measures.