Sakana Fugu Cyber

Fugu Cyber, released on 21 July 2026, applies Sakana's orchestration idea to security work: a pool of specialised agents coordinated behind a single endpoint that behaves like one model. Sakana reports 86.9% on CyberGym, which asks an agent to read complex codebases and confirm real vulnerabilities, and 72.1% on CTI-REALM, which asks it to turn raw threat-intelligence reports into working detection rules — figures the company calls comparable to security-tuned frontier models such as GPT-5.5-Cyber and Mythos Preview. What makes the announcement unusual is that half of it argues against its own product category: Sakana writes plainly that scoring well on an evaluation is only the start, that raw models deployed alone will produce false positives, and that anything a model flags has to pass specialised sub-agents and human review before a patch is proposed. The company frames the model as one piece of enterprise defence rather than a solution, and says its Applied Enterprise team is building the surrounding workflows with large Japanese institutions. Access is deliberately narrow: the model sits on the token (pay-as-you-go) plan only, is covered by an updated acceptable-use policy forbidding offensive use, and every applicant must file a request describing the intended use case with verified contact details, reviewed manually before a key is issued. Sakana did not publish parameter counts or the composition of the agent pool; the context window of one million tokens comes from the vendor's own API documentation.

▸Videos

No videos yet.

▸Reviews