Microsoft has announced 'MAI-Cyber-1-Flash,' a security-focused AI that, when combined with GPT-5.4, delivers security capabilities exceeding Claude Mythos 5 at a low cost.
Microsoft has announced that it has integrated its AI security model, ' MAI-Cyber-1-Flash ,' into its AI security system, '
Introducing MAI-Cyber-1-Flash inside MDASH | Microsoft AI
https://microsoft.ai/news/introducing-mai-cyber-1-flash-inside-mdash/
'MDASH' is an agent-based vulnerability detection and remediation system announced by Microsoft in May 2026. It is described as a system that uses multiple AI models to scan for bugs that could be exploited.
Microsoft points out that with the remarkable advancements in AI, cyber threats are also intensifying, and the security model of 'scan occasionally and patch eventually' is now outdated. They explain that they developed MAI-Cyber-1-Flash because they believe that building a superior model is necessary to help enhance the robustness of software running worldwide in order to unlock the true benefits of AI.
MAI-Cyber-1-Flash is designed to identify hard-to-find vulnerabilities within complex codebases, is deeply integrated with MDASH, and has been proven robust in the world's largest security environments.
The following are the results from CyberGym, an industry-standard benchmark that evaluates a system's reasoning ability to analyze large codebases and discover vulnerabilities within them. Combining MAI-Cyber-1-Flash with GPT-5.4 outperformed 'Gemini 3.5 Flash Cyber,' 'GPT-5.5 Cyber,' 'GPT-5.6 Sol,' and 'Claude Mythos 5.'
According to Microsoft, security is a mission that requires constant operation, and with a massive influx of attacks, token costs are the biggest obstacle for defenders.
Since MAI-Cyber-1-Flash is designed to efficiently handle up to 90% of all tasks, MDASH can leverage the largest and most costly model in its model family (in this case, GPT-5.4) for the extremely difficult 10% of tasks that are truly needed.
The MDASH integrated system equipped with MAI-Cyber-1-Flash achieved a score of 96% in CyberGym, and its cost can be reduced by 50% compared to the best MDASH configuration to date (GPT-5.4 + 5.4mini + 5.3codex).
While AI-based fixes for software vulnerabilities have become a crucial part of the workflow, many tasks still remain for human personnel. Therefore, MDASH has released 'Perception,' an agent-based security system that provides an agent team to support various security workflows within MDASH, performing continuous monitoring, patching, and containment of new threat vectors.
Perception plans to utilize MAI-Cyber-1-Flash in many more security workflows in the future, not just for addressing software vulnerabilities.
Related Posts:
