GLM-5.3 is here with advanced cyber capabilities — and reportedly already found a 'serious vulnerability' in Cursor

Chinese AI startup Z.ai released GLM-5.3, reporting major gains in long-horizon coding and cybersecurity from expanded post-training rather than a new base model. The company says the model has already identified a potentially serious Cursor vulnerability and will delay API access and open weights until safety evaluation and hardening are complete.

GLM-5.3 highlights how much frontier models can improve through post-training alone, reducing the need for costly new pretraining runs. Its rapid jump in offensive cyber capability also underscores the growing dual-use challenge for open-model developers as they balance openness with access controls.