Microsoft has unveiled a new suite of AI-driven security tools, arguing that the same agentic AI capable of finding software flaws for attackers can be turned into a faster, cheaper defender. At an event on Monday, the company introduced its first security-specific model, MAI-Cyber-1-Flash, alongside an agent-coordination system called Project Perception, positioning both as central to its strategy for keeping pace with AI-accelerated cyberattacks.

Microsoft says MAI-Cyber-1-Flash, built on its internal MAI-Thinking-1 reasoning model and paired with GPT-5.4 inside its MDASH bug-hunting harness, scored 95.95 per cent on the CyberGym vulnerability benchmark, beating rivals including OpenAI's GPT-5.5 Cyber (85.6 per cent) and GPT-5.6 Sol (83.6 per cent), Anthropic's Mythos 5 (83.8 per cent) and Google's Gemini 3.5 Flash Cyber (83.2 per cent), while costing roughly half as much. The smaller model handles about 90 per cent of queries itself, passing the rest to the larger GPT-5.4. Microsoft also launched Project Perception, which uses red, blue and green team agents to find, assess and fix vulnerabilities, plus a new research division called FORGE Labs and an "External Red Team Alliance" offering unrestricted funding to 18 university AI safety labs worldwide.