GLM-5.3: AI Finds 1,097 Critical Bugs Post-Training

Z.ai's new GLM-5.3 model has found 1,097 critical and high-severity vulnerabilities in real software. What's interesting is that the model's cybersecurity ability grew far beyond what the company intended during training. This unexpected capability led to a delayed open-weight release for safety review. The company states that all performance gains in GLM-5.3 came from post-training, not from changes to the base model or pretraining. The same base model from GLM-5.2 was used, with Z.ai running the training stack on more diverse environments for a longer time. These environments are designed to mimic professional work. This shows how advanced AI models can develop unforeseen capabilities, impacting cybersecurity.