Z.ai GLM-5.3 reaches 84.5% on CyberGym vulnerability benchmark LLM performance via agentic fine-tuning, surpassing GLM-5.2 and proprietary rivals while triggering safety delays.