Sakana AI's Fugu-Cyber: New Cybersecurity Model Excels

Sakana AI has introduced Fugu-Cyber, a new cybersecurity model. This model achieved an 86.9% success rate on CyberGym and 72.1% on CTI-REALM benchmarks. These results are comparable to other top cybersecurity models. Fugu-Cyber is part of Sakana's Fugu orchestration family and is specifically tuned for security reasoning. CyberGym measures the ability to write proof-of-concept exploits for real-world vulnerabilities. CTI-REALM evaluates the mapping of threats and creation of detection rules. For CyberGym, Sakana's 86.9% is a slight improvement over previously reported top scores. On CTI-REALM, Fugu-Cyber's 72.1% places it above previous top configurations, though this benchmark uses a different scoring method. The Fugu orchestrator uses a language model to assign tasks to specialized models. This approach involves roles like Thinker, Worker, and Verifier for different language models. For security tasks, the verifier role is key, validating vulnerabilities before any patch suggestions. This development could impact how cybersecurity threats are identified and addressed.