Microsoft Unveils First Specialized AI Security Model For Cyber Defense
Its argument is that conventional security tooling, built around the assumption of human attackers, can no longer keep pace. Autonomous systems now generate exploits faster, scale campaigns further, and run around the clock while the cost of launching an attack keeps falling. That same rise in AI-driven risk has already pushed lawmakers toward emergency kill switch powers after a major AI breach, and Project Perception aims to counter the shift by continuously analyzing risk across an organization's identities, endpoints, applications, data, cloud services, and AI systems.
The system leans on three groups of specialized AI agents that work as a closed loop. Red team agents probe an environment for paths to compromise before real attackers find them, blue team agents investigate those findings and sort out which ones present meaningful risk, and green team agents handle remediation and harden defenses. Microsoft says human defenders still set the objectives and guardrails, with high-impact actions requiring sign-off from a person rather than firing automatically.
Rather than lean on one large model for every job, Project Perception uses a multi-model architecture that picks models based on quality, reliability, speed, and cost, mixing frontier models with smaller ones trained for specific cybersecurity tasks. The first of those purpose-built models is MAI-Cyber-1-Flash, now integrated into MDASH, the company's multi-model team of software vulnerability agents. Microsoft says the updated configuration scored 96% on CyberGym, an industry benchmark for evaluating AI on vulnerability work, landing 12 percentage points ahead of Mythos while cutting costs by nearly half compared to the existing MDASH setup.
Those agents do not start each investigation from scratch. Project Perception maintains a continuously updated picture of an organization's assets, identities, relationships, past incidents, policy decisions, and active risks, then hands every agent access to that shared context. The payoff should be less time and computing power spent gathering and reconstructing information from raw security signals before any real work begins.
Microsoft also plans to wire Project Perception directly into its security products. The system arrives first inside Microsoft Defender, where agents turn findings into corrective actions rather than dumping yet another pile of alerts on already stretched security teams, with support expanding across the broader portfolio over time.
Project Perception enters public preview on August 3 with consumption-based pricing measured in Security Compute Units, and more demanding agent tasks will burn through more of them. The bet is that pairing specialized models with a coordinated team of autonomous defenders can help organizations keep up as attacks grow faster, cheaper, and increasingly hands-off.
