Sakana AI's Fugu-Cyber Matches GPT-5.5 on the Hardest Security Benchmarks

Fugu-Cyber is Sakana AI's new cybersecurity-specialized variant of its Fugu orchestration model. It hits 86.9% on CyberGym and 72.1% on CTI-REALM -- two of the hardest real-world security benchmarks available -- putting it shoulder-to-shoulder with dedicated cyber models like GPT-5.5-Cyber and Mythos Preview. But the more interesting part of the release is not the benchmark numbers. It is the argument Sakana is making about what those numbers actually mean in production.

What Fugu Is, and Why Cyber Is Different

To understand Fugu-Cyber, you need to understand what Fugu is. Fugu is itself a language model specialized to understand when to delegate, how agents should communicate, and how to combine their work into a single, reliable answer -- an approach built on Sakana's ICLR 2026 research on learned model orchestration, including the TRINITY and Conductor papers. In plain terms: instead of you wiring together a LangChain graph and deciding which model calls which, Fugu learns that coordination itself.

TRINITY is a lightweight evolved coordinator that assigns Thinker, Worker, and Verifier roles across a multi-model pool and adaptively delegates across coding, math, and reasoning tasks. The Conductor is trained with reinforcement learning to discover natural-language coordination strategies -- effectively learning how a diverse set of models should talk to each other so the group outperforms any single worker. TRINITY is a roughly 0.6B-parameter coordinator, evolved with CMA-ES, that assigns these roles across a pool of much larger worker models.

Fugu-Cyber takes that same architecture and specializes it for security workflows. Like the original Fugu, Fugu-Cyber is a multi-agent system that behaves like a single model -- you send a request to one endpoint, and the system dynamically orchestrates a pool of specialized agents to tackle complex, multi-step tasks without the risk of single-vendor dependency.

The Benchmarks: What They Actually Test

The two benchmarks Fugu-Cyber targets are worth understanding in detail, because they are meaningfully harder than most cyber evals.

• CyberGym -- a large-scale benchmark featuring 1,507 real-world vulnerabilities across 188 software projects, primarily tasking agents with generating a proof-of-concept test that reproduces a vulnerability, given only its text description and the corresponding codebase. Fugu-Cyber scores 86.9% here.