Chinese Open-Source AI Solution DoGNAVY Ranks Third Globally in CyberGym Benchmark

The Chinese AI safety solution DoGNAVY, co-developed by DarkNavy, ranked third globally and first among open-source models in the CyberGym AI safety benchmark with a 90.8% pass rate. It achieved this milestone using only a single open-source model, GLM-5.2, whereas higher-ranking competitors relied on ensembles of closed-source models. This achievement demonstrates that open-source models can compete with proprietary, closed-source ensembles in complex cybersecurity tasks. It lowers the barrier to entry for advanced AI safety capabilities, allowing more innovators to deploy top-tier security agents locally. While Microsoft and Google routed tasks dynamically across multiple top-tier closed-source models, DoGNAVY relied entirely on GLM-5.2, a model designed for long-horizon tasks and coding. CyberGym evaluates AI agents across a dynamic, real-world lifecycle of vulnerability discovery, proof, and patching.