Microsoft Launches MAI-Cyber-1-Flash: MDASH System Reaches 95.95% on CyberGym at Half the Previous Cost

My take: On July 27, Microsoft introduced MAI-Cyber-1-Flash, its own specialized cybersecurity model. Integrated into the MDASH system, which handles 90% of tasks while GPT-5.4 covers the most complex 10%, the combined setup reached 95.95% on the CyberGym benchmark, twelve points above Anthropic Mythos (83.8%), its nearest competitor. The operating cost, according to Microsoft, is half that of its previous configuration.

The numbers are solid, but worth reading with a critical eye: Microsoft is both judge and party, since the benchmark was run by the same company that makes the product. It is worth waiting for independent evaluations before treating these figures as definitive, especially given that recent research has shown that CyberGym scores reported by different vendors are not always directly comparable.

What is clear is the market direction. AI tools for cybersecurity are already a present reality, and competition among Microsoft, Anthropic, and others is producing more capable specialized models that, according to their makers, are also more cost-accessible. The system enters public preview on August 3 in Azure AI Foundry.

Is your security team already evaluating or using AI to detect and close vulnerabilities, or does it still depend exclusively on manual processes?