Microsoft Perception helps security teams identify high-risk AI agents, understand their exposure and coordinate actions to reduce risk. It is aimed at security operations and administrators responsible for agent governance across cloud and endpoint environments. Microsoft Events demonstrates agent discovery, risk investigation and remediation workflows in a Microsoft Ignite session.
Agents can change risk as they receive new permissions, connect to more data or tools, and take on additional autonomous actions. Perception is designed to give security teams a continuously updated view of that agent estate rather than treating an inventory or an alert as the whole picture. Its agent-security experience connects discovery, investigation and defensive actions.
In the demonstration, Perception brings cloud agents and local agents into an AI estate view, alongside tools and MCP servers. It uses signals from Microsoft security products to add context, including identity information, permissions and risk assessments. The system can explain why an agent is flagged, show relevant access and display an attack-surface graph that the session identifies as shared with Microsoft Defender.
Microsoft describes the wider approach through red, blue and green security agents. Red agents probe for weaknesses, blue agents detect and investigate threats, and green agents work on fixes and posture improvements. Playbooks connect findings to a desired outcome: assess the estate, identify gaps, draft or recommend a fix, and, when authorized, deploy controls or trigger security actions.
Perception is aimed at security teams that need to discover agents, prioritize their risk and coordinate protection across an organization. The session describes visibility into cloud and local agents, and says that connecting additional services can expand what the system discovers. In its example, Defender helps reveal local agents, while Global Secure Access is suggested to help observe network traffic and MCP servers.
The demonstration also describes connecting to Microsoft Agent 365 to see non-Microsoft platforms, including Salesforce Agentforce connections. That makes agent visibility relevant to organizations using a mixture of Microsoft and third-party tools. The system can incorporate information from Microsoft Entra, Defender, Purview and other security controls to build shared context.
Microsoft Perception was described as generally available at Ignite 2026. The agent-estate and remediation capabilities were presented as extensions to that system. Separately, Microsoft lists the Posture Prioritization Agent as generally available and the Risk Reduction Agent as in preview; those stages apply to those named exposure-management agents.
The demonstrated workflow starts with an inventory and uses playbooks to add context and guide action. Security teams can investigate a high-risk result before deciding whether to make a change. The steps shown can be summarized as follows:
The session's example agent had been marked compromised by an administrator in the Entra admin center and had a Microsoft Graph permission that could create agents. Perception surfaced those details to explain why the agent's risk mattered. This illustrates how identity status and permissions can help distinguish an urgent case from an agent that is simply present in the inventory.
Discovery depends on the sensors and connections available to Perception. The demonstration shows that deploying Defender exposes local agents, while additional connectivity can help uncover network traffic, MCP servers and agents on non-Microsoft platforms. Teams should therefore review which parts of their estate are connected before treating an inventory as complete.
Automated remediation also has a governance boundary. The session says green agents can deploy controls if given permission, and emphasizes that security teams remain in control of actions. Playbooks can recommend work and coordinate it, but teams should review the proposed outcome and authorization before allowing a consequential change.
Availability also differs by capability: Perception is generally available, while the Risk Reduction Agent is in preview. The Posture Prioritization Agent is generally available and separates prioritization from execution, allowing organizations to apply their own approval processes. That distinction matters when deciding whether a capability can recommend a response or is ready for a production workflow.
It helps security teams discover agents, assess risk, investigate alerts and coordinate controls or remediation through playbooks.
It combines available security context, such as agent identity, permissions and risk signals, to flag agents and explain why they matter.
Microsoft described Perception as generally available at Ignite 2026. The Risk Reduction Agent is in preview, while the Posture Prioritization Agent is generally available.
Green agents can deploy controls and trigger actions when authorized. Security teams retain control over whether to permit those actions.