Z.ai has unveiled GLM-5.3, the latest iteration of its AI model, designed to tackle complex coding tasks and bolster cybersecurity analysis. This version builds upon the foundation of GLM-5.2, with enhancements achieved through extensive post-training at a larger scale.
Unlike its predecessor, GLM-5.3 emphasizes training AI agents within realistic task environments rather than isolated coding exercises. These environments encompass comprehensive codebases, documentation, storage systems, experimental results, testing tools, and intricate multi-step workflows. For instance, an agent might be tasked with identifying a bottleneck in a machine learning training stack, implementing an optimization, conducting tests, and demonstrating improved performance without compromising functionality.
The model has demonstrated significant improvements in coding benchmarks. On Terminal-Bench 3.0, GLM-5.3 achieved a score of 28.3, a substantial increase from GLM-5.2’s 4.6. Similarly, on DeepSWE v1.1, the score rose to 66.9 from 46.2. Internal evaluations using Z.ai’s Code Bench indicated a 50% enhancement in performance within realistic local development settings. GLM-5.3 offers three reasoning settings: low, high, and max, with the max setting recommended for coding tasks to allow the model to dedicate more effort to planning, implementation, testing, and verification. Notably, this version does not support the complete disabling of reasoning.
Cybersecurity capabilities have also seen notable advancements. By incorporating vulnerability discovery data and security-focused task environments during post-training, GLM-5.3 has improved in identifying and analyzing vulnerabilities. On the CyberGym benchmark, which assesses white-box vulnerability discovery in source code, the model scored 84.5%, up from 77.2% in the previous version. In ExploitBench, its score increased from 24.4% to 54.4%. Additionally, in ExploitGym, GLM-5.3 completed 105 exploitation tasks within two hours and 130 within six hours, compared to GLM-5.2’s 29 and 39 tasks, respectively.
These enhancements suggest that GLM-5.3 is more adept at connecting various stages of an attack path, including vulnerability analysis and exploitation reasoning. This capability could assist defenders in identifying complex weaknesses involving multiple components and chained flaws. However, it’s acknowledged that some leading closed models still outperform GLM-5.3 on certain exploitation benchmarks.
In practical applications, GLM-5.3 has been tested against real-world codebases, identifying 2,436 vulnerabilities across 269 projects after expert review and duplicate removal. Of these, 1,097 were rated medium to high severity, affecting software such as kernels, operating systems, browser engines, web applications, network protocols, and open-source infrastructure. To manage these findings, Z.ai has established a public Security Disclosure Ledger to track coordinated disclosures. At launch, 53 findings had been publicly disclosed, with 2,383 under embargo. Notably, the oldest reported flaw dates back to 1981, highlighting the longevity of hidden vulnerabilities in widely used code. Model weights are expected to be released two weeks post-launch, following safety evaluations and hardening.
The release of GLM-5.3 marks a significant step forward in AI’s role in coding and cybersecurity. By enhancing the model’s ability to operate within realistic environments and improving its vulnerability detection capabilities, Z.ai is contributing to more robust and efficient security practices. As AI continues to evolve, such developments are crucial in addressing the increasingly complex challenges in software development and cybersecurity.
