Chinese AI company Zhipu claims its new is a better bug-finder than Anthropic, OpenAI

Chinese AI company Zhipu recently unveiled its new AI model, GLM-5.3, which it claims to be as proficient at finding bugs as the leading American models, such as Anthropic's Mythos and OpenAI's GPT-5.6 Sol. The company released benchmark data showing GLM-5.3 surpassing both Fable 5 and GPT-5.6 Sol in the CyberGym benchmark, which assesses a model's ability to address cybersecurity challenges in real-world scenarios.

Zhipu's statement highlights that the model's cyber capabilities developed at an unprecedented pace during post-training, outperforming expectations. The company reported that GLM-5.3 is currently the state-of-the-art model in vulnerability discovery on the CyberGym benchmark, with the greatest improvements occurring as the model advances through the exploitation stages and forms coherent plans for complete exploitation chains.

After testing the model on real-world codebases, Zhipu discovered 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues. These vulnerabilities were found in various components, ranging from system kernels and operating systems to browser engines, open-source infrastructure, web applications, and network protocols.

While GLM-5.3 performed exceptionally well in bug detection, the model underperformed when compared to western models in other security and coding benchmarks. Despite this, the fact that China has developed such a highly-capable bug finder demonstrates that the country is rapidly catching up in terms of cybersecurity prowess. This development undermines any advantage the US once had as the home of major AI companies like Anthropic.