Sakana AI releases Fugu-Cyber, a security-tuned model

Sakana AI released Fugu-Cyber, an orchestration model built on its Fugu system and tuned for offensive/defensive cybersecurity tasks. It reports 86.9% on the CyberGym benchmark and 72.1% on CTI-REALM, edging out GPT-5.5-Cyber and Claude Mythos Preview on those tests. Access is gated behind manual approval, a defensive-use policy, and Sakana's Token Plan.

Why it matters: Purpose-built cybersecurity models are becoming a distinct competitive lane alongside general-purpose LLMs, following similar cyber-focused releases from OpenAI and Anthropic. The gated access and defensive-use policy reflect the industry's ongoing struggle to balance offensive-security research value against misuse risk, a tension also visible in recent reports of AI models being used in real breaches.