Today, CyberGym, an internationally recognized AI security benchmark, announced the latest ranking. DogNavy, which was jointly developed by top domestic artificial intelligence research institutes and security team Daku Novi, ranked third in the world and number one in open source with a 90.8% pass rate. CyberGym is currently the largest cybersecurity test for AI agents. Why can the Chinese team rank among the top three? Two technical routes gave the answer. Microsoft and Google, which are at the top, rely on a combination of top closed source models to “fight”, and each step routes the best performing model — provided they can buy and use the world's strongest closed source model at the same time. DogNavy follows a different path: using only an open source generic model, anyone can download and freely deploy. Notably, the Chinese solution DogNavy only used an open source generic model, while the top rivals relied on a combination of multiple closed source models. AI security attacks and defense are entering the practical stage, and the open source route is making top security capabilities mastered by more innovators.

Today, CyberGym, an internationally recognized AI security benchmark, announced the latest ranking. DogNavy, which was jointly developed by top domestic artificial intelligence research institutes and security team Daku Novi, ranked third in the world and number one in open source with a 90.8% pass rate. CyberGym is currently the largest cybersecurity test for AI agents. Why can the Chinese team rank among the top three? Two technical routes gave the answer. Microsoft and Google, which are at the top, rely on a combination of top closed source models to “fight”, and each step routes the best performing model — provided they can buy and use the world's strongest closed source model at the same time. DogNavy follows a different path: using only an open source generic model, anyone can download and freely deploy. Notably, the Chinese solution DogNavy only used an open source generic model, while the top rivals relied on a combination of multiple closed source models. AI security attacks and defense are entering the practical stage, and the open source route is making top security capabilities mastered by more innovators.