Chinese AI GLM-5.3: Advanced Bug-Finding Capabilities

A new Chinese AI model, GLM-5.3, claims to find bugs as well as leading American models. It showed strong performance on the CyberGym benchmark, a test for real-world cybersecurity challenges. The model reportedly outperforms Fable 5 and GPT-5.6 Sol on this benchmark, identifying vulnerabilities and building exploitation chains. Zhipu tested GLM-5.3 on real codebases, finding 2,436 vulnerabilities across 269 projects. Over a thousand of these were medium to high severity, spanning system kernels, operating systems, browser engines, and network protocols. Some vulnerabilities were reportedly 40 years old. While GLM-5.3 performed less effectively on other benchmarks, its rapid progress in vulnerability discovery suggests a closing gap in AI-driven cybersecurity. This development could impact future global cybersecurity efforts.