Microsoft's Project Perception AI Security Platform Enters Preview

Microsoft has opened Project Perception to public preview, a security platform that puts coordinated AI agents to work defending networks against AI-driven attacks. The company detailed the rollout in a July 27 blog post and confirmed the system went live for outside testing on August 3.

Three agents, one loop

Project Perception organizes its work across three agent roles. Red-team agents simulate attacks to map how an organization could be compromised. Blue-team agents investigate what the red team finds and reason about which of those paths represent real risk. Green-team agents then act on the findings, from closing misconfigurations to hardening exposed systems. An orchestrator routes tasks between the three, and Microsoft says the system keeps a full audit trail of what each agent did and why.

The company says human approval is still required before an agent can take a “consequential” action, such as changing a production system, and that permissions are scoped narrowly rather than granted broadly.

Built on a mix of models

Rather than one general model, Project Perception runs on a combination of frontier and purpose-built systems, chief among them MAI-Cyber-1-Flash, a cybersecurity-specific model Microsoft introduced alongside the platform. Microsoft says the model scores 96% on CyberGym, an industry benchmark for finding and reproducing software vulnerabilities — 12 points above Anthropic’s Claude Mythos, a rival cybersecurity model — while running at roughly half the cost of Microsoft’s earlier configuration for the same task.

Limited rollout so far

At launch, Project Perception is integrated only into Microsoft Defender, though Microsoft says it plans to extend the system across its broader security product line over time. Pricing is consumption-based, billed in units the company calls Security Compute Units, so cost rises with how much agent activity a customer’s environment generates. Microsoft has not disclosed a general-availability date, regional rollout plan, or which customer tiers can access the preview.

Microsoft frames the launch around a broader shift in cybersecurity, where attackers increasingly use AI to move faster than human defenders can respond, arguing that countering AI-driven threats requires AI systems capable of acting at similar speed.