Sakana AI Unveils Fugu‑Cyber: Multi‑Agent AI for Network Defense

Sakana AI, a Tokyo‑based AI model developer, announced on July 21, 2026 the launch of Fugu‑Cyber, a multi‑agent AI orchestration model designed specifically for cybersecurity operations.

Fugu‑Cyber integrates several specialized AI agents that collaborate autonomously after a user sends a command to a single API endpoint. This coordination enables the system to handle multi‑stage security workflows such as large code‑base analysis, vulnerability verification, threat‑intelligence analysis, and detection‑rule generation.

According to the company’s blog and disclosed test data, Fugu‑Cyber achieved world‑class performance on key security benchmarks. In the “CyberGym” test, which evaluates real‑world vulnerability analysis using 1,507 vulnerability data points from 188 software projects, the model attained an 86.9% success rate. In the “CTI‑REALM” benchmark, measuring end‑to‑end detection‑rule generation, it reached a 72.1% success rate.

These results are reported to be comparable to cutting‑edge, security‑focused models such as OpenAI’s GPT‑5.5‑Cyber and Anthropic’s Claude Mythos Preview.

The article notes the dual‑use challenge in cybersecurity AI: traditional large language models often trigger security filters that block legitimate defensive actions, while overly relaxed controls increase attacker risk. Fugu‑Cyber addresses this by dynamically coordinating multiple dedicated agents, providing a secure environment for verification tasks without compromising safety.

Access to Fugu‑Cyber is currently invitation‑only and subject to a strict vetting process. The service operates under a new usage policy that prohibits abusive behavior and is offered only through enterprise‑grade, pay‑as‑you‑go packages. Prospective users must submit a detailed form describing their intended use and verified contact information for pre‑approval.