Sakana AI's Fugu-Cyber Matches GPT-5.5 on the Hardest Security Benchmarks
Sakana AI's Fugu-Cyber matches GPT-5.5-Cyber on security benchmarks by orchestrating specialist agents -- but argues raw scores are only half the story
- Fugu-Cyber launches: Sakana AI releases a cybersecurity-specialized orchestration model scoring 86.9% on CyberGym and 72.1% on CTI-REALM, matching GPT-5.5-Cyber and Mythos Preview.
- Multi-agent under the hood: Like base Fugu, it dynamically orchestrates a pool of specialist agents behind a single OpenAI-compatible API endpoint -- no manual workflow wiring required.
- Gated access: Available only on the pay-as-you-go Token Plan ($6/$36 per million input/output tokens); users must apply and be manually approved before gaining access.
- Reality check on cyber AI hype: Sakana argues raw model access does not solve enterprise security -- false positives, lack of codebase context, and missing human-in-the-loop verification are the real blockers.
- Enterprise focus: Sakana's Applied Enterprise team is building specialized deployment harnesses for major Japanese institutions, combining Fugu-Cyber with human security expertise.
- Research-backed architecture: Built on two ICLR 2026 papers (TRINITY and Conductor) on learned model orchestration -- the coordinator itself is trained, not hand-wired.
Fugu-Cyber is Sakana AI's cybersecurity-specialized variant of its Fugu orchestration model. It scores 86.9% on CyberGym and 72.1% on CTI-REALM, two of the hardest real-world security benchmarks available, putting it alongside dedicated cyber models like GPT-5.5-Cyber and Mythos Preview. The benchmark numbers are strong, but the more interesting part of the release is the argument Sakana makes about what those numbers actually mean in production.
What Fugu Does, and Why Cyber Is Different
Fugu is a language model trained to coordinate other models: when to delegate, how agents should communicate, and how to combine their outputs into a single reliable answer. The architecture comes from Sakana's ICLR 2026 research, specifically the TRINITY and Conductor papers. Where a typical LangChain setup requires you to wire together a graph and decide which model calls which, Fugu learns that coordination directly.
Two research threads underpin it:
- TRINITY is a roughly 0.6B-parameter coordinator, evolved with CMA-ES, that assigns Thinker, Worker, and Verifier roles across a pool of larger worker models, adapting dynamically across coding, math, and reasoning tasks.
- Conductor is trained with reinforcement learning to discover natural-language coordination strategies, effectively teaching a diverse model pool how to communicate so the group outperforms any individual member.
Fugu-Cyber applies that same architecture to security workflows. You send a request to one endpoint; the system routes it across a pool of specialized agents and returns a single answer, without requiring you to manage the orchestration or accept single-vendor dependency.
The Benchmarks: What They Actually Test
Both benchmarks are meaningfully harder than most cyber evals, and worth understanding in detail.
- CyberGym covers 1,507 real-world vulnerabilities across 188 software projects. Agents receive a vulnerability's text description and the corresponding codebase, then must generate a proof-of-concept test that reproduces it. Fugu-Cyber scores 86.9%.