Microsoft announced Project Perception, an agentic security system designed by the company to address an environment in which the speed, scale, and economics of attacks are being reshaped by artificial intelligence. The system enters public preview on August 3, offering an approach aimed at turning security signals into real-time protection while keeping human defenders in control.
Microsoft says security methods built for a world dependent on human actors can no longer keep pace with systems that can reason, adapt, and operate continuously, alongside attacks carried out at machine speed. Project Perception therefore relies on a system that continuously monitors risks across the digital environment, understands broad context, identifies appropriate actions, and learns from changes in the environment.
Three Types of Agents in a Continuous Defense Loop
Project Perception combines signals, context, models, and specialized agents operating within a continuously learning system. The platform coordinates three types of agents:
- Red team agents: They identify potential paths for compromising the environment before attackers exploit them.
- Blue team agents: They investigate signals, analyze context, and identify significant risks.
- Green team agents: They carry out corrective actions and strengthen protections across the environment.
Together, these agents form a closed loop that continuously detects, assesses, and improves an organization's security posture, rather than merely generating more alerts.
A Multilayered, Multimodel Security Architecture
The system begins with signals and sensors that provide visibility across digital assets. Security context then transforms these signals into an understanding that agents can use efficiently. Models provide reasoning capabilities, while the control component coordinates models and agents within security workflows, and execution components turn decisions into protective actions.
Microsoft says Project Perception provides continuously updated security context that connects data, knowledge, and semantics related to assets, identities, relationships, risks, and activities. This allows agents to directly access the information needed to analyze risks and prioritize actions, rather than recollecting and linking context from raw signals each time.
The platform also uses a multimodel architecture that selects the most suitable model for each task based on quality, reliability, response time, and cost. As part of this approach, Microsoft introduced the MAI-Cyber-1-Flash model into MDASH, a multimodel team of agents for software vulnerability management.
According to the company, the MDASH configuration with MAI-Cyber-1-Flash achieved a score of 96% on the CyberGym benchmark, 12 points ahead of Mythos, while also achieving approximately 50% cost savings compared with the current MDASH configuration available on the market. Microsoft plans to use the same model in additional security workflows within Project Perception, while continuing to develop specialized models.
Broad Visibility and Actionable Measures
The platform draws on Microsoft's visibility across identities, endpoints, applications, data, clouds, and artificial intelligence systems, as well as its expertise in security research, threat intelligence, and real-world operations. It also integrates with Microsoft Security products, allowing agents to connect insights to actions and reduce risks, rather than merely detecting them.
Microsoft emphasizes that Project Perception was built according to its responsible AI principles and inherits the security, compliance, governance, and operational controls on which customers rely. The system will be available to customers worldwide through public preview starting August 3.