Microsoft has introduced MAI-Cyber-1-Flash, its first specialized cybersecurity AI model, as part of a multi-model approach to software vulnerability management inside MDASH. The company positions the model within Project Perception, a broader defense system that coordinates specialized agents and enters public preview on August 3 for customers around the world.
Microsoft says MDASH with MAI-Cyber-1-Flash scores 96% on the CyberGym benchmark, 12 points above Mythos. The same configuration delivers almost 50% cost savings versus the current MDASH setup in market.
The company frames that result as the product of a tuned multi-model system with access to rich historical training data.
Project Perception is built around continuous understanding of attacker paths, defender risk evaluation, and ongoing posture improvement. Red team agents surface potential compromise paths, blue team agents investigate context and meaningful risk, and green team agents apply corrective actions.
Microsoft states that effective cyber defense requires applying the right model to the right problem at the right time rather than relying on a single frontier model for every task.
The stack also includes security context that turns broad visibility across identities, endpoints, applications, data, clouds, and AI systems into shared, near real-time understanding for agents. Microsoft says the architecture optimizes for quality, reliability, latency, and cost so protection stays highly effective, continuously available, and affordable at scale.
Next steps include expanding MAI-Cyber-1-Flash beyond software vulnerability workflows into additional security tasks inside Project Perception.
