Z.ai Delays GLM-5.3 Open Weights Two Weeks After Post-Training Produced Unplanned Exploit-Chain Reasoning
Z.ai released GLM-5.3 on August 14, 2026 on the same base model as GLM-5.2, attributing a claimed 50% coding gain entirely to scaled post-training, and ranking first among open-source models on Terminal Bench 3.0 and Agents' Last Exam. On CyberGym it scored 84.5%, up from GLM-5.2's 77.2% and slightly ahead of GPT-5.6 Sol, but the company says vulnerability-discovery training environments produced an unintended emergent ability to reason across full multi-stage exploitation chains rather than isolated bugs. Weights are withheld roughly two weeks for safety hardening — reportedly the first time a Chinese lab has cited emergent offensive capability as the reason; Z.ai says its models have found 2,436 vulnerabilities across 269 open-source projects since GLM-5.2, 1,097 rated critical or high.
↳ Follow the thread