China's Z.ai holds GLM-5.3 weights for two weeks over exploit-chaining risk

Z.ai shipped GLM-5.3 on the same 743B-parameter base as GLM-5.2, taking its gains from scaled post-training alone — 84.5 on CyberGym, ahead of Claude Fable 5 at 83.8 and GPT-5.6 Sol at 83.6, with Terminal-Bench 3.0 jumping from 4.6 to 28.3. The open weights are held back roughly two weeks for safety hardening after the model began chaining multi-stage exploits, with only selected security partners granted controlled access in the interim.

出典： axios.com ↗