Sakana AI builds a cyberdefense model from a swarm of agents

Frontier models are getting scarier at finding software bugs, but enterprise security teams know the dirty secret: drop an unharnessed LLM into production, and it mostly just drowns your engineers in false alarms.

Tokyo-based Sakana AI is tackling that gap head-on. The lab just released Fugu-Cyber, a multi-agent orchestration model built specifically for enterprise cyber defense, accessible via a dedicated API endpoint.

On raw benchmarks, Fugu-Cyber trades blows with the biggest names in the game. It hit 86.9% on CyberGym, which tests how well an AI can analyze codebases and confirm real-world vulnerabilities. On CTI-REALM, a test that measures turning raw threat intel into working detection rules, it scored 72.1%. Those numbers put it on par with specialized frontier models like GPT-5.5-Cyber and Anthropic's Mythos-Preview.

A multi-agent squad behind a single API

Fugu-Cyber looks and feels like a standard model endpoint, but under the hood, it is an orchestration engine. Send it a complex defensive task, and it dynamically delegates steps across a pool of specialized agents, preventing teams from being locked into a single model vendor.

Sakana also took a rare swipe at frontier model hype. Citing a recent report from Nikkei Digital Governance, the lab pointed out that even Japan's biggest financial giants struggle to use raw frontier models like Mythos. Without deep access to proprietary codebases and human verification pipelines, raw models simply fail to deliver reliable patches.

To bridge that divide, Sakana's Applied Enterprise team is working directly with major Japanese institutions. They are building verification harnesses where specialized sub-agents and human security pros validate an alert in an isolated staging environment before any fix goes anywhere near production.

Because of the offensive risks involved, access is tightly gated. Beyond rolling out a stricter Acceptable Usage Policy that bans offensive attacks, Sakana is requiring teams to submit detailed use cases for manual human vetting before granting API access on its Token Plan.

Why it matters

We think Sakana's approach is a realistic reality check for the entire security industry.

Most enterprise AI pitches pretend a single frontier model can replace a security operations center overnight. In reality, a tool that hallucinates a fake exploit or pushes a broken patch can be just as damaging as an actual breach. By framing Fugu-Cyber as an orchestrator that keeps humans and multi-model fallbacks in the loop, Sakana is selling operational reliability rather than sci-fi promises.

For enterprise security leads, this is worth watching closely, though it will not be a plug-and-play fix for small teams. The manual vetting process and need for deep internal integration mean the early beneficiaries will be banks, utilities, and large enterprises.

There is also a clear geopolitical angle. Sakana explicitly linked this release to AI sovereignty. For organizations that want frontier-grade defenses without tying their digital security entirely to a single American hyperscaler, an orchestrator that coordinates multiple models behind localized walls looks like a very clever hedge.