The internationally recognized AI security benchmark CyberGym announced its latest rankings today. DoGNAVY, jointly developed by leading Chinese artificial intelligence research institutions and the security team DARKNAVY, ranks third globally and first among open-source systems with a 90.8% pass rate.
According to CCTV News, CyberGym is currently the largest-scale cybersecurity test for AI agents. Why has the Chinese team managed to break into the top three? Two technical approaches provide the answer.
Microsoft and Google, which rank ahead, rely on combinations of multiple top-tier closed-source models, routing each step to the best-performing model — on the premise that the world's most powerful closed-source models can be purchased and used simultaneously.
DoGNAVY takes a different approach: it uses only one open-source general-purpose model, GLM-5.2, released as open source by Zhipu in June, which anyone can download and deploy freely.
It is worth noting that the Chinese solution DoGNAVY uses only one open-source general-purpose model, while the higher-ranked competitors rely on combinations of multiple closed-source models.
CCTV News stated that AI security offense and defense are entering a practical stage, and the open-source approach is enabling more innovators to gain access to top-tier security capabilities.