GLM-5.3: Frontier Coding with Emergent Cyber Capabilities
Z.ai says its GLM-5.3 model gained substantially on long-horizon coding and vulnerability-analysis tasks through post-training, while also developing unexpectedly strong exploitation capability.
The company reports a 50% improvement over GLM-5.2 on its private coding benchmark, 84.5% on CyberGym, and more than twice GLM-5.2’s score on ExploitBench. It also claims 2,436 findings across 269 open-source projects, though only 53 are publicly disclosed and the results are vendor-run, so independent validation matters. Z.ai plans to release the model weights after safety evaluation and hardening.