Microsoft is preparing to test a new approach to enterprise defense with Project Perception, an agentic security system intended to continuously discover, evaluate and reduce cyber risk. Rather than presenting another assistant that summarizes alerts, Microsoft is positioning the technology as an operational layer capable of coordinating security decisions and actions across identities, endpoints, applications, data and cloud environments.
Three Types of Agents
Project Perception divides its work among three classes of specialized agents. Red team agents search for potential attack paths before adversaries can use them. Blue team agents investigate activity and determine which findings represent meaningful risk. Green team agents apply corrective actions and strengthen the environment.
These agents operate with access to a shared representation of an organization's assets, identities, relationships and security conditions. This context is intended to reduce the time and computing resources agents would otherwise spend reconstructing the environment from raw alerts.
A Multi-Model Security Architecture
Microsoft is also avoiding dependence on a single AI model. Project Perception uses a multi-model architecture that can select a model according to the required quality, latency, reliability and cost. Its initial vulnerability-management scenario incorporates MAI-Cyber-1-Flash into MDASH, Microsoft's team of software vulnerability agents.
The company says this configuration achieved a 96 percent result on the CyberGym benchmark while reducing costs by almost half compared with the current MDASH configuration. Those results should still be evaluated carefully under real enterprise conditions, where incomplete inventories, unusual software and conflicting business requirements can complicate automated remediation.
Control Must Remain Visible
In my view, the most important part of Project Perception is not its ability to generate findings. It is the proposed connection between analysis and controlled action. Security teams already receive more alerts than they can comfortably process. An agentic platform becomes useful only when it can prioritize accurately, document its reasoning and make reversible changes under clear human supervision.
Organizations evaluating the August 3 public preview should begin with limited permissions and isolated workflows. Every agent action should be logged, high-impact changes should require approval, and rollback procedures should be tested before production use. Project Perception reflects where security platforms are heading, but its long-term value will depend on whether machine-speed response can be delivered without sacrificing accountability.