Sakana AI Launches Fugu-Cyber with High Benchmark Scores Amid Scrutiny

• •Fugu-Cyber claims an 86.9% score on CyberGym and 72.1% on CTI-REALM benchmarks.
• •The benchmark methodology has not been disclosed, raising questions about the validity of the scores.
• •Sakana AI emphasizes the need for specialized human expertise in cybersecurity for effective deployment.

On July 21, 2026, Sakana AI introduced Fugu-Cyber, a cybersecurity orchestration system that claims an 86.9% score on the CyberGym benchmark and 72.1% on CTI-REALM. These scores, if validated, would position Fugu-Cyber above competitors like OpenAI's GPT-5.5-Cyber. However, the methodology behind these benchmarks has not been disclosed, raising concerns about the validity of the claims. The Fugu-Cyber system operates as a multi-agent architecture, allowing it to dynamically assign tasks to specialized agents for complex cybersecurity tasks. Sakana AI emphasizes that achieving high benchmark scores does not equate to solving real-world security challenges, which require specialized human expertise and integration into existing systems. The company warns against over-reliance on frontier models without proper operational support. Security engineers are advised to critically evaluate the benchmarks and the system's deployment requirements before use.