Microsoft just dropped its first in-house cybersecurity AI model, MAI-Cyber-1-Flash – and the numbers are turning heads.
The 5-billion-active-parameter model is built specifically for finding tricky vulnerabilities in complex code. It’s integrated into MDASH, Microsoft’s multi-agent vulnerability hunting system, and handles roughly 90% of security tasks on its own. Only the hardest 10% get kicked up to OpenAI’s GPT-5.4.
The result? A 95.95% score on CyberGym – the industry benchmark for vulnerability detection. That beats Anthropic’s Mythos 5 (83.8%), GPT-5.5 Cyber (85.6%), and Google’s offerings. Microsoft says that’s a 12-point jump over Mythos.
But here’s the kicker: it costs 50% less than Microsoft’s previous MDASH setup (GPT-5.4 + GPT-5.4 mini + GPT-5.3 Codex). Mustafa Suleyman, CEO of Microsoft AI, put it bluntly: “We have world-leading performance at 50% of the cost”.
The secret sauce is routing – matching each task to the right model instead of throwing everything at one expensive frontier model. It’s a bet that the future isn’t about the biggest model, but the cheapest one that’s good enough, used intelligently.
Microsoft also unveiled Project Perception – an agentic security system with red, blue, and green team AI agents that hunt threats, assess risk, and fix defenses in a continuous loop. It enters public preview on August 3.
The bigger picture? Microsoft’s betting on its data moat – decades of security signals across identity, endpoint, cloud, and network. As Suleyman put it: “This is genuinely the tip of the iceberg”.
Microsoft’s MAI-Cyber-1-Flash is available inside MDASH, not as a standalone public model." I write this blog, now I need a minimalist, ultra-clean banner image for this technology news blog. Try to avoid text as it’s looks scattered
Tags:
Technology