Microsoft Project Perception Enters Public Preview: What Security Teams Should Know

Microsoft's Project Perception has entered public preview, introducing coordinated AI agents into security operations. This move raises concerns about permissions, oversight, and deployment risks. The system, announced on Aug. 3, features specialized agents designed to investigate threats, assess risk, and implement defensive actions.

Project Perception extends Microsoft's AI-driven analysis to coordinated security operations, working across various security data and tools. However, the broader authority of these agents raises immediate questions regarding permissions, accuracy, oversight, and auditability.

The system uses a Cyber Stack model, integrating security signals, organizational context, multiple AI models, and specialized agents. Three classes of agents are involved: red-team agents identify potential compromise paths, blue-team agents investigate activity, and green-team agents take corrective actions to enhance defenses. Microsoft ensures human operators retain control, though customers must define approval criteria and narrow each agent's permissions.

Recent AI agent permission issues highlight the need for careful consideration. Microsoft's model performance report indicates that its system scored 95.95% on the CyberGym vulnerability benchmark, using MAI-Cyber-1-Flash for most tasks and GPT-5.4 for the most complex ones. The configuration costs nearly 50% less than previous setups. Despite these benefits, organizations should start with narrowly scoped permissions and maintain human oversight for actions that could impact protections or production systems.

Security teams should evaluate the tools and data each agent consumes, as poisoned MCP tool descriptions can lead to unintended actions. Useful pilot metrics include investigation accuracy, false-positive rates, analyst review time, and the percentage of recommendations requiring correction. Teams must also ensure that agent inputs, decisions, approvals, and outcomes are fully auditable.

While Microsoft's true public pricing and general-availability date remain undisclosed, the preview should be viewed as a controlled evaluation rather than a basis for immediate production deployment.