Microsoft sets August public preview for Project Perception, an agentic security system

Microsoft has outlined Project Perception, an agentic security system that it says will enter public preview on August 3. The company describes the project as a new security stack for an environment in which attackers and defenders can both use autonomous systems. Rather than treating AI as an add-on to alerting tools, Microsoft is presenting Perception as a system intended to continuously gather signals, reason over context and take protective actions while people remain in control.

At the center of the design are three groups of specialized agents. Red-team agents are meant to identify possible paths to compromise before an attacker uses them, blue-team agents investigate and prioritize risk, and green-team agents take corrective action and strengthen defenses. Microsoft describes the combination as a closed loop: discover a weakness, assess its significance and improve the security posture rather than leaving a human team with another isolated alert.

The company is placing unusual emphasis on shared security context. Project Perception is designed to connect information about assets, identities, applications, data, cloud services, AI systems, relationships and activity into a continuously updated picture that agents can use. The aim is to reduce the repeated effort of collecting and correlating raw signals before a security decision can be made, while giving operators a common basis for reviewing what the system proposes.

Microsoft is also making a multi-model argument. It says no single model is suitable for every security task, so Perception will combine frontier and specialized cyber models according to quality, reliability, latency and cost. The first named scenario is software vulnerability management, where the company plans to use MAI-Cyber-1-Flash inside its MDASH multi-model team of agents. Microsoft says that configuration reached 96% on CyberGym and reduced costs relative to its current configuration, claims that will need to be assessed as the preview reaches users.

The proposal is consequential because autonomous security systems increase both the speed of response and the importance of oversight. Microsoft says the project is integrated across its security products and built on existing security, compliance, governance and operational controls. In practice, customers will need to evaluate what actions agents can take automatically, how their reasoning is reviewed and how teams can reverse or contain a decision when the context is incomplete.

The August preview is therefore the beginning of a product test rather than the final proof of the architecture. Microsoft will need to show that the shared context is accurate across complex estates, that the agent roles produce useful outcomes instead of more noise, and that the economics of an always-on multi-model service hold up in production. Those details will determine whether Project Perception becomes a practical operating layer for defenders or remains chiefly a statement of direction.