Microsoft has opened the public preview of Project Perception, an agentic cybersecurity system designed to help enterprises identify, evaluate and remediate software vulnerabilities at machine speed.
The preview began Aug. 3, 2026, following Microsoft’s July 27 introduction of the platform and MAI-Cyber-1-Flash, its first internally trained cybersecurity model. Project Perception combines specialized AI agents, security context and multiple models in a system intended to move beyond generating alerts and toward taking coordinated defensive action.
Three agent roles form a security loop
Project Perception divides work among red, blue and green agents. Red agents look for potential paths an attacker could exploit. Blue agents investigate the findings, add context and determine which issues represent meaningful risk. Green agents take corrective action and strengthen defenses.
Microsoft describes the three roles as a closed loop that continuously discovers, evaluates and improves an organization’s security posture. The system draws on signals across identities, endpoints, applications, data, cloud infrastructure and AI systems, then provides agents with a shared view of assets, relationships and risk.
Human control remains part of Microsoft’s stated design. That qualifier is important. An agent that recommends a patch presents a different operational risk from one authorized to change a production system, particularly when a fix could interrupt revenue, customer access or critical workflows.
Microsoft is optimizing for speed and cost
The platform uses a multi-model architecture rather than assigning every task to one large model. Its first workflow places MAI-Cyber-1-Flash inside MDASH, Microsoft’s multi-model system for software vulnerability analysis, while reserving other models for work requiring different capabilities.
Microsoft says the configuration scored 96% on CyberGym, a benchmark for producing working proofs of concept for known vulnerabilities, and reduced costs by nearly 50% compared with the company’s previous MDASH configuration. Those performance and cost figures are Microsoft’s own. TechRadar noted that outside researchers had not independently verified the claims at publication.
The economic premise is consequential for enterprise buyers. Continuous scanning and remediation can become expensive when every decision is routed through a frontier model. Using smaller specialized models for common tasks and more capable models for difficult cases could make always-on defense more practical, provided accuracy, oversight and rollback controls meet enterprise requirements.
Automation raises the governance stakes
Project Perception reflects a broader shift in cybersecurity: AI is moving from an assistant that summarizes information to an operator that can initiate changes. That transition may help defenders respond faster, but it also expands the decisions organizations must govern. Axios reported that defenders remain cautious about handing autonomous agents security authority.
CIOs, CISOs and boards evaluating agentic security should define which actions agents may take independently, which require human approval and how every intervention will be logged, tested and reversed. They should also determine who is accountable when automated remediation creates downtime or fails to address the underlying risk.
Microsoft says Project Perception follows its responsible AI principles and inherits existing security, compliance and operational controls. For enterprise leaders, those assurances are a starting point rather than a substitute for internal policy and testing.
The strategic value of agentic defense will not be measured by the number of vulnerabilities an AI system can surface. It will be measured by whether organizations can convert machine-speed detection into reliable action without surrendering operational control.