GLM-5.3: The Open-Weights Model That Got Scary Good at Coding and Cybersecurity Overnight

Z.ai unveiled GLM-5.3, an open-weights AI model that displays remarkable improvements in coding and cybersecurity. The model achieved a 50% boost on coding benchmarks, outperformed state-of-the-art performance on multiple agentic benchmarks, and developed surprising emergent cybersecurity capabilities. GLM-5.3 uses the same base model as GLM-5.2, with enhancements coming from post-training techniques like reinforcement learning and fine-tuning.

Post-training is proving to be a significant factor in AI model performance. Z.ai scaled three key aspects of their training: more diverse tasks, more complex environments mimicking real-world work, and increased computational power. These scaled environments enabled GLM-5.3 to develop advanced cybersecurity reasoning skills.

Z.ai introduced vulnerability discovery data into GLM-5.3's training mix, focusing on identifying and reasoning about multiple stages of exploitation. As a result, the model showed significant improvements in identifying and exploiting vulnerabilities. It excelled in vulnerability identification (84.5% accuracy) and exploitation reasoning (54.4%), outperforming previous models in both areas.

The real-world impact of GLM-5.3 is evident in its ability to find 2,436 vulnerabilities across 269 open-source projects, with 1,097 being medium-to-high severity issues. Many of these vulnerabilities had remained unnoticed for years. Z.ai has launched a public Security Disclosure Ledger to track these findings as they progress through the disclosure process.

The advancements in GLM-5.3 demonstrate that the future of AI improvement lies in more sophisticated post-training techniques rather than larger pre-training runs. This open-weights model has matched or surpassed closed models on specific capabilities, such as the CyberGym SOTA.