Sakana AI

Cybersecurity · AI Orchestration · Enterprise AI

Launch

Sakana AI launches Fugu-Cyber, scoring 86.9% on CyberGym benchmark

The orchestrator routes tasks across specialist models, letting enterprises avoid depending on any single AI vendor for cyber defense.

• Sakana AI released Fugu-Cyber, a new API endpoint version of its Fugu orchestration model, hitting 86.9% on the CyberGym benchmark and 72.1% on CTI-REALM.
• Fugu is a multi-agent system that presents a single API but internally routes subtasks to a pool of specialized models, verifying and synthesizing their outputs into one response.
• CyberGym tests an agent's ability to analyze complex codebases and verify real-world vulnerabilities, while CTI-REALM measures turning raw threat intel reports into working detection rules.
• Fugu first launched June 22, 2026 as a general-purpose orchestrator benchmarked against Anthropic's Fable 5 and Mythos Preview, and it has since added Nvidia's Nemotron models as coding and tool-use specialists.
• Sakana AI pushed back on 'fearmongering' about frontier cyber models, arguing that raw capability alone doesn't solve enterprise security without deep integration and human expertise.
• As Mythos-class cyber models proliferate from Anthropic, OpenAI, and Chinese labs, Sakana AI is betting that orchestrating many specialist models can match frontier performance without single-vendor lock-in.