# Microsoft turns Project Perception into a closed-loop security system for the AI era Source: TechNewsList (https://technewslist.com) Canonical URL: https://technewslist.com/en/article/microsoft-project-perception-agentic-security-2026-08-03-morning Section: AI (https://technewslist.com/en/ai) Author: TechNewsList Language: en Published: 2026-08-03T05:10:01.716+00:00 Updated: 2026-08-03T05:10:01.895306+00:00 > Microsoft is putting Project Perception into public preview with red, blue, and green agents that continuously find, investigate, and remediate security risk while leaving consequential decisions under human control. ## TL;DR - Microsoft announced that Project Perception enters public preview on August 3, 2026. - The system coordinates red, blue, and green agents to discover vulnerabilities, investigate risk, and apply corrective action. - Its security context spans identities, endpoints, applications, data, clouds, and AI systems. - Microsoft is using a multi-model architecture and says its first vulnerability-management scenario cuts model cost by almost half. - The product is designed to automate continuous defense while people retain control over high-impact decisions. ## Key points - Project Perception treats security as a continuous loop of perception, reasoning, and action rather than a queue of alerts. - Red, blue, and green agents divide offensive discovery, defensive investigation, and remediation work. - Shared security context reduces the need for every agent to rebuild the organization’s state from raw signals. - The multi-model approach matches specialist and frontier models to individual security tasks. - Usage-based Security Compute Units make the economics of always-on agentic defense part of the product design. # Microsoft turns Project Perception into a closed-loop security system for the AI era Security products have spent years turning machine activity into alerts for human teams to triage. Microsoft’s Project Perception is built around a more ambitious loop: continuously understand an organization, reason about changing risk, and take bounded action before an attacker’s path becomes an incident. The system enters public preview on August 3, 2026. ![Microsoft Project Perception agentic security system graphic.](https://cdn-dynmedia-1.microsoft.com/is/image/microsoftcorp/4252354-Hero-SCIAgenticSystem?fit=constrain&hei=563&op_usm=1.5%2C0.65%2C15%2C0&qlt=100&resMode=sharp2&wid=1000) ## What happened Microsoft announced Project Perception in a July 27 security post as an agentic system designed for an environment where both defenders and attackers can operate continuously. Rather than adding an AI assistant to an existing dashboard, Microsoft describes a new stack that combines signals, shared security context, models, an orchestration harness, agents, and actuators that can turn a decision into a protection. The system is organized around three specialized roles. Red team agents look for potential compromise paths before an attacker can exploit them. Blue team agents investigate signals, connect them to organizational context, and decide what represents meaningful risk. Green team agents take corrective actions and strengthen protections. The handoff between those roles is the product’s central design choice: discovery should feed investigation, and investigation should feed remediation without requiring a person to manually restart the workflow at every stage. ![Diagram of Project Perception’s coordinated security agents.](https://cdn-dynmedia-1.microsoft.com/is/image/microsoftcorp/4252354-OverviewMultiAgent-SCIAgenticSystem?fit=constrain&hei=1283&op_usm=1.5%2C0.65%2C15%2C0&qlt=100&resMode=sharp2&wid=2000) Microsoft says the first scenario is software vulnerability management through MDASH, a multi-model team of agents. The company also says a configuration using its MAI-Cyber-1-Flash model reached 96% on CyberGym, twelve points above Mythos, while cutting model cost by almost half versus the current MDASH configuration. Those figures are vendor-reported benchmarks, but they show the direction: security agents must be evaluated on useful work and sustainable economics, not only on how fluent their explanations sound. ## Why it matters The hard problem in agentic security is not generating a plausible recommendation. It is maintaining enough context to know whether that recommendation applies, then connecting it to a safe action. Microsoft says Project Perception can observe identities, endpoints, applications, data, clouds, and AI systems. That breadth lets the system build a near-real-time representation of assets, relationships, risks, and activity instead of asking each agent to reconstruct the environment from raw telemetry. That shared context is important because an isolated model can mistake an unusual action for an attack, miss a relationship between two accounts, or propose a fix that breaks a production dependency. A system that knows the organization’s previous incidents, policy decisions, and active environment can reason with more evidence. It can also expose the context behind a conclusion, which is essential when a human defender needs to approve or reject a high-impact action. The design also reflects a practical limit of single-model systems. Security work ranges from fast classification to deep investigation, code analysis, threat hunting, and remediation planning. Microsoft says Project Perception continuously chooses models using quality, reliability, latency, and cost as criteria. That is closer to a software architecture decision than a model leaderboard contest. ## Technical details Microsoft frames the product as a stack rather than a single agent. Sensors and signals provide awareness. Security context turns those signals into usable organizational meaning. A model layer supplies reasoning. A harness coordinates agents and models. Agents apply the reasoning to security workflows, while actuators connect decisions to changes in the environment. The distinction between insight and action matters. A chatbot can explain a vulnerability; an actuator can open a ticket, change a control, or isolate a resource. Project Perception is designed to support the latter, but Microsoft says people remain in control of important decisions. The product documentation describes human-set strategy, scoped permissions, traceability, and approval for consequential actions. Microsoft’s product page also describes consumption-based pricing through Security Compute Units. Different agents consume those units at different rates depending on task intensity. That pricing model makes operational efficiency part of the security architecture: an always-on defense system has to be affordable enough to run continuously, not only during a crisis. ![Project Perception red, blue, and green agent roles.](https://cdn-dynmedia-1.microsoft.com/is/image/microsoftcorp/project-perception-red-blue-green-agents?fit=constrain&hei=540&op_usm=1.5%2C0.65%2C15%2C0&qlt=100&resMode=sharp2&wid=960) ## Market / industry impact Project Perception puts pressure on the security market’s existing division of labor. Endpoint, identity, cloud, application, and AI-security tools have historically produced separate findings that teams correlate in a security operations center. A coordinated agent layer could make the value shift from collecting more alerts to maintaining better context and executing safer workflows. It could also change how organizations buy models. If a security platform can route each task to a specialist model and keep the decision inside a governed workflow, customers may care less about choosing one universal model. They will care more about evaluations, audit trails, data boundaries, rollback, and the ability to prove why an action happened. The risk is that a closed loop can close too quickly. A mistaken red-team finding could trigger an incorrect blue-team judgment, which could prompt a green-team remediation that causes an outage. Human control therefore has to be operational, not a marketing phrase: permissions, approvals, explanations, reversibility, and monitoring must be built into the system. ## What to watch next The public preview will show whether the product can move beyond impressive demonstrations into repeatable vulnerability-management work. Watch the quality of its evidence, the rate of false positives, the boundaries around green-team actions, and how customers audit model choices. The broader AI signal is that enterprise security is becoming a systems problem. The winning security agent will not be the one that talks most convincingly. It will be the one that maintains context, knows when to ask for approval, and turns a defensible decision into a measurable reduction in risk. ## Sources - [Microsoft Security Blog](https://blogs.microsoft.com/blog/2026/07/27/rethinking-security-for-the-age-of-ai/) - July 27, 2026 announcement. - [Project Perception product page](https://www.microsoft.com/en-us/security/business/ai-powered-cybersecurity/project-perception-agentic-system) - Agent roles, controls, and product architecture. - [Microsoft Security](https://www.microsoft.com/en-us/security/business/ai-powered-cybersecurity/project-perception-agentic-system) - Official operational and governance context. Category signal: ai. Mentions: Microsoft, Project Perception, Microsoft Defender, MAI-Cyber-1-Flash, agentic security, vulnerability management, Security Compute Units ## Sources - [Microsoft Security Blog](https://blogs.microsoft.com/blog/2026/07/27/rethinking-security-for-the-age-of-ai/) - [Microsoft Project Perception](https://www.microsoft.com/en-us/security/business/ai-powered-cybersecurity/project-perception-agentic-system) - [Microsoft AI](https://www.microsoft.com/en-us/security/business/ai-powered-cybersecurity/project-perception-agentic-system)