On August 6th, the international authority CyberGym, which tests the cybersecurity benchmark of AI agents, released the latest assessment list. The AI security agent DoGNAVY, jointly developed by the domestic artificial intelligence research institute and the Dacunovy security team, ranked third globally with a 90.8% test passing rate, while at the same time securing first place in comprehensive results for open-source models, highlighting our country‘s breakthrough progress on the AI security autonomous technology path.
CyberGym is currently the world‘s most widely covered and most recognized AI intelligent body cybersecurity practical assessment system, and the assessment results have industry benchmark reference value. In this ranking, the top-ranked overseas products generally adopt multiple sets of top closed-source large model fusion scheduling technologies, improving security attack and defense performance through the procurement of multiple commercial closed-source models and task-separated optimal invocation methods. This model has limitations such as high deployment costs, high usage thresholds, and technology system non-autonomy.
In stark contrast, the domestic DoGNAVY has taken a completely differentiated, autonomous, open-source technology path. This agent completes full-process security attack and defense inference solely based on a single open-source universal large model spectrum GLM-5.2. The relevant code and model resources are open to the entire society, allowing various enterprises, scientific research units, and technology developers to download them for free and deploy them independently, significantly reducing the threshold for using top-notch AI security capabilities.
Industry insiders say that the current global AI security offensive and defense has fully entered the stage of practical development. For a long time, high-end AI security capabilities have been highly concentrated in closed-source commercial systems, with core technology and computing power resources in the hands of a few overseas technology companies. The domestic open-source AI security solution, represented by DoGNAVY, breaks down the inherent perception that multi-closed-source model stacking can achieve top-notch offensive and defense capabilities, proving that based on local open-source large models, it is also possible to build AI security intelligences that reach the level of international first-tier teams. (Chun Jun)