Microsoft launches compact cybersecurity model MAI-Cyber-1-Flash

Microsoft introduced MAI-Cyber-1-Flash, a compact security model that scores 96% on the CyberGym benchmark when run inside its MDASH multi-agent system. Microsoft says the setup cuts costs roughly 50% versus using a pure frontier model, since only the hardest cases get escalated to GPT-5.4.

Why it matters: Despite building its own security model, Microsoft still routes its toughest cases to OpenAI's GPT-5.4, underscoring that even well-resourced competitors aren't ready to replace frontier-model reasoning with smaller in-house models. It also lands just after 30+ companies, notably excluding OpenAI, Google, and Anthropic, formed a rival open-source security alliance, highlighting a real split in how the industry wants to handle AI-driven cybersecurity.