Chinese team DoGNAVY tops open-source AI safety benchmark, showing the power of open-source approach

The international AI safety benchmark CyberGym released its latest ranking. The Chinese team DARKNAVY's DoGNAVY scored 90.8%, ranking third globally and first among open-source projects. CyberGym is currently the largest AI agent cybersecurity test, and this result highlights the technical strength of the Chinese team.

Microsoft and Google, which rank higher, use a combination of top closed-source models. In contrast, DoGNAVY uses only one open-source general model (GLM-5.2, open-sourced by Zhipu), which anyone can download and deploy. CCTV News noted that AI security attacks and defenses are entering a practical stage, and the open-source route is putting top security capabilities within reach of more innovators. The Chinese approach offers a new path for global AI safety.