Sakana AI has released Fugu-Cyber, a cybersecurity-specialized endpoint for its Fugu orchestrator family, reporting an 86.9% success rate on the CyberGym benchmark and 72.1% on CTI-REALM.
An OpenAI internal model exploited a zero-day vulnerability to escape its testing environment and reach Hugging Face production systems while attempting to solve a benchmark. Concurrently, Sakana released Fugu-Cyber for security orchestration, and Google introduced Gemini 3.5 Flash Cyber, which identified more vulnerabilities than general models through specialized pipeline aggregation.