Fugu-Cyber

Fugu-Cyber is Sakana AI's orchestration model tuned specifically for cybersecurity work — vulnerability discovery, exploit reasoning, and threat-intelligence analysis — released July 21, 2026 as a new API endpoint. Where general frontier models handle security tasks as one capability among many, Fugu-Cyber is trained and evaluated against domain benchmarks, and Sakana positions its scores on CyberGym (real-world vulnerability tasks) and CTI-REALM (cyber threat intelligence) as state-of-the-art, matching the dedicated cyber variants of GPT-5.5.

It's built for security engineers, red teams, and threat researchers who want a model that behaves like a coordinated agent across security tooling rather than a chat window. As an "orchestration" model, the pitch is that it plans and sequences multi-step security workflows — enumerating a target, reasoning about a vulnerability class, drafting analysis — instead of answering one prompt at a time. This is a raw model accessed through an API, not a packaged app, so you're wiring it into your own pipeline or agent framework.

The honest caveat: nearly everything known about performance comes from Sakana's own release, and access is gated behind an application-and-approval process with no free tier or public playground. That makes independent verification hard right now, and the specialization cuts both ways — this is not a general-purpose model you'd reach for outside security contexts.