Japan’s Sakana AI releases Fugu Cyber: a multi-agent system that combines GPT-5

Japan's Sakana AI releases Fugu Cyber: a multi-agent system that defeats both GPT-5.5-Cyber ​​and Claude. The line of defense of network security is being redefined by a new combat unit. On July 21, Japanese artificial intelligence start-up Sakana AI released the Fugu Cyber ​​model, which is specially designed to deal with the complexity of modern network defense. It has delivered results that are enough to outperform its closed-source opponents in the industry's most challenging security benchmark tests. Specifically, Fugu Cyber ​​achieved a success rate of 86.9% in the CyberGym test and a success rate of 72.1% in the CTI-REALM test. The performance was even better than OpenAI’s GPT-5.5-Cyber ​​and Anthropic’s Claude Mythos-Preview. This means that in the face of real and tricky attack and defense scenarios, the startup’s dedicated model has already run ahead of the general flagship. Fugu Cyber's capabilities come from its system form. In line with the standard version of Fugu, it is a multi-agent system encapsulated into a single API. For each user request, the system will dynamically orchestrate different professional agents to collaboratively handle complex multi-step tasks, rather than relying on a single model. In addition to this simple set of APIs, Sakana AI also provides on-site deployment services for network security solutions supported by enterprise application teams, bringing model capabilities directly into customers' production environments. via AI News (author: AI Base)