Sakana AI's Fugu-Cyber Just Beat GPT-5.5 on Cyber Benchmarks — Then Issued a Reality Check

The Spine: Japan's Nature-Inspired AI Lab Just Proved You Don't Need a Trillion-Parameter Model to Win at Cybersecurity

In the high-stakes world of AI-powered cyber defense, the conventional wisdom has been simple: bigger models win. GPT-5.5-Cyber. Claude Mythos. The trillion-parameter club.

Sakana AI just threw a wrench into that orthodoxy. The Tokyo-based startup, whose name means "fish" in Japanese and whose research draws inspiration from natural evolutionary processes, released Fugu-Cyber — an orchestration model that matches frontier systems on enterprise security benchmarks without training a single large foundation model of its own.

The numbers are striking: 86.9% on CyberGym (a benchmark testing vulnerability verification in complex codebases) and 72.1% on CTI-REALM (evaluating the ability to turn raw threat intelligence into working detection rules). These scores put Fugu-Cyber on par with OpenAI's GPT-5.5-Cyber and Anthropic's freshly-reapproved Claude Mythos-Preview.

But the more interesting story isn't the benchmark score. It's the reality check Sakana attached to it.

A Multi-Agent System, Not a Mega-Model

Fugu-Cyber isn't a single large language model. It's an orchestration layer — a system that sits on top of a pool of specialized sub-agents and dynamically routes tasks to the right tool for each step of a security workflow.

"Like our original Fugu orchestration model, Fugu-Cyber is a multi-agent system that behaves like a single model," Sakana explained in its blog post. "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."

This architectural choice is core to Sakana's philosophy: instead of competing on scale (building bigger transformer models with more parameters and more training compute), the company competes on orchestration — stitching together smaller, specialized models in ways that produce frontier-competitive results.

The approach is particularly well-suited to cybersecurity, where workflows are inherently multi-step: detect an anomaly, correlate it with threat intelligence, analyze the affected codebase, verify the vulnerability, and recommend a remediation — each step potentially requiring a different specialized capability.

The Reality Check Sakana Wants You to Read

Buried deeper in the announcement is a section titled "The Reality Check on Frontier Cyber Capabilities" — and it's here that Sakana makes its most provocative argument.

"Achieving high scores on an evaluation is only the beginning of the story," the company writes. "Recently, there has been a lot of fearmongering about the cyber capabilities of frontier models. Much of the industry narrative suggests that simply granting an organization access to a frontier model with cyber capabilities will instantly solve their security challenges."

Sakana doesn't mince words: "We believe it is time to ground this conversation in reality."

Citing a recent Nikkei Digital Governance report (in Japanese), the company notes that large organizations — "including major financial institutions" — often struggle to operationalize even the most powerful frontier cyber models. Without specialized internal talent and deep integration into proprietary source code and infrastructure, a GPT-5.5-Cyber or Mythos access pass doesn't magically transform an enterprise's security posture.

This is a self-aware message from a company that just shipped a frontier-competitive cyber product: the tool alone isn't enough. The ecosystem around it — integration, talent, processes — matters just as much.

Access Gated Behind Manual Approval

Fugu-Cyber is available now as a new API endpoint at sakana.ai/fugu, but getting access requires jumping through a few hoops. The company gates access behind manual approval, a defensive-use policy, and the Token Plan (Sakana's usage pricing system).

This mirrors the access restrictions placed on GPT-5.5-Cyber and Anthropic's Mythos — a recognition that powerful cyber capabilities, even in defensive hands, carry dual-use risks. A model that can analyze codebases for vulnerabilities can also be used to find exploitable holes. The defensive-use policy attempts to ensure the tool stays in defenders' hands.

Not Everyone Is Convinced

Tech Times raised questions about the lack of disclosed benchmark methodology in Fugu-Cyber's announcement, noting that independent verification of the 86.9% and 72.1% scores is not yet available. Sakana has not published the full evaluation setup, dataset splits, or reproducibility details that would allow third-party validation — a common issue in the rapidly-moving AI benchmarking landscape where companies race to claim state-of-the-art results.

The comparison to GPT-5.5-Cyber also warrants scrutiny. OpenAI's cyber model was designed for red-teaming and offensive security testing, while Fugu-Cyber's benchmarks focus on defensive enterprise workflows (vulnerability verification and detection rule generation). The two models serve partially overlapping but distinct use cases, and head-to-head comparisons should account for these differences.

What This Means for the AI-Security Landscape

Sakana's approach represents a genuine third path in AI development. While American labs race to scale (OpenAI, Anthropic, Google) and Chinese labs race to catch up (Alibaba's Qwen, DeepSeek), Sakana is betting that orchestration — not scale — wins in specialized domains.

If Fugu-Cyber's independent benchmarks hold up, the implications extend far beyond cybersecurity. The same multi-agent orchestration approach could be applied to legal analysis, medical diagnosis, financial modeling, or any domain where complex workflows demand the coordination of multiple specialized capabilities rather than raw general intelligence.

Japan's AI ecosystem has long been underestimated in the global conversation dominated by US and Chinese labs. Sakana — backed by investors including Lux Capital and backed by a team of former Google Brain researchers — is making a case that the country's AI future doesn't depend on outspending its rivals, but on outthinking them.

Sources

• Sakana AI: Fugu-Cyber Release (Official Blog Post)
• MarkTechPost: Sakana AI Releases Fugu-Cyber
• Tech Times: Sakana AI Fugu-Cyber Claims 86.9% Vulnerability Score; Benchmark Methodology Not Disclosed
• AIBase: Japan's Sakana AI Releases Fugu Cyber: A Multi-Agent System That Overshadows GPT-5.5-Cyber and Claude
• Sakana AI: Fugu 1.1 with Claude Code Interface
• Nikkei Digital Governance Report (referenced by Sakana, in Japanese)