Zhipu's GLM-5.3 Tops US Rivals in CyberGym Test

Zhipu, a Chinese artificial intelligence company also referred to as Z.ai, has introduced its premier model, GLM-5.3, asserting that it surpassed top American systems in cybersecurity evaluations. The Beijing-based firm reported that GLM-5.3 recorded an 84.5% success rate on CyberGym, a standard for assessing a model's capability to detect and confirm security vulnerabilities in source code. This result exceeded Anthropic's Mythos 5, which achieved 83.8%, and OpenAI's GPT-5.6 Sol at 83.6%. However, GLM-5.3 only managed a score of 54.4% on ExploitBench, falling short of Mythos's 78% and GPT-5.6 Sol's 76.5%. Zhipu noted that its model was tested with security teams in China, uncovering 2,436 vulnerabilities across 269 projects, with 1,097 classified as medium to high severity. This launch aligns with China's initiative to bolster its capabilities in AI-driven cybersecurity. A representative from Zhipu highlighted the importance of global cooperation in AI development rather than competition among nations.