HomeServicesManaged IT ServicesAI DeploymentAccredOneInsightsAnswersAboutFree Consultation
AI Security & Governance

Microsoft AI Agent Security: Build the Control Plane Before Scale

Microsoft is turning AI agents into first-class security objects. With Microsoft Agent 365 and Defender, security teams can build an inventory, assess posture, inspect runtime behavior, block supported high-risk actions, and investigate agent activity alongside identities, devices, applications, and incidents.

The change is operational as well as technical. Effective July 1, 2026, Microsoft moved the listed security capabilities for Copilot Studio and Microsoft Foundry agents behind an Agent 365-eligible license. Tenants that previously relied on Defender for Cloud Apps or Defender for Cloud alone need to confirm licensing and recheck their enablement.

The key distinction: Agent 365 is the control plane and registry; Defender is the security operations layer. Neither automatically makes every agent, tool, permission, or execution path safe.

Oct. 8, 2026 update: Microsoft Security’s Oct. 7 platform guidance makes the boundary explicit: trust is an end-to-end property of the agent system, not a feature you add after the model works. The rest of this guide keeps the Defender detail below, but expands the operating model to cover the agent’s identity, data, tools, memory, runtime, and evidence.

What the new platform guidance changes

Microsoft’s security model now treats an agent as a system that can receive untrusted content, reason over context, call tools, access enterprise data, retain memory, and act at machine speed. Prompt injection is one risk, but it is not the whole risk. Excessive permissions, unsafe connectors, weak tool contracts, stale memory, unowned identities, and unobservable runtime paths can all turn a useful agent into a high-impact incident.

For administrators and security leaders, the practical shift is to place controls at the points where an agent receives input, retrieves data, chooses a tool, and executes an action. Logging an outcome after the fact is useful evidence; it is not a substitute for an enforceable boundary before a sensitive tool call.

Control plane questionMinimum evidence to keepOwner
Who is this agent and who is accountable?Unique identity, owner, sponsor, business purpose, platform, lifecycle status, and decommission date.IT governance and the business process owner
What can it read and change?Data sources, sensitivity, delegated or autonomous permissions, tools, APIs, MCP/A2A connections, and approval requirements.Security architecture and data owners
What happens when input or behavior is hostile?Prompt-injection tests, tool-call policy, runtime detections, audit events, rollback steps, and the person who can pause or revoke the agent.Security operations and the agent owner
Can the organization prove what happened?Authentication, data-access, tool-invocation, policy, approval, and response records with defined retention and privacy access.Security, compliance, and records management

Map the agent before you secure it

Start with a one-page system map rather than a product checklist. Record the interaction channel, model and orchestration layer, tools and connectors, retrieval sources, short- and long-term memory, identity and access controls, runtime environment, and security operations path. This map makes unsupported or unmonitored routes visible before a pilot becomes production infrastructure.

  • Interactive agents: document the user context and delegated permissions they inherit.
  • Autonomous agents: give each one a distinct identity and constrain its permissions to the job, not the platform’s maximum capability.
  • Agent user accounts: treat persistent mailboxes, Teams access, and document permissions as high-impact assets that need an owner and expiry path.
  • Agent-to-agent workflows: authenticate both sides, restrict which agents can delegate, and retain an auditable chain of calls.

Microsoft Entra Agent ID provides the identity and lifecycle foundation for these patterns. Where Agent 365 is not yet available for a scenario, Microsoft’s Cloud Adoption Framework still calls for a centralized inventory, a distinct identity, ownership, data governance, security monitoring, and cost visibility using the services already in the tenant.

Oct. 29, 2026 update: Microsoft announced general availability for local AI agent discovery and configured MCP server visibility. Microsoft Defender for Endpoint Plan 2 is the minimum license for inventory and hunting. One scope detail needs care: Microsoft’s announcement says Windows and macOS, while the current Learn support matrix still labels macOS discovery as preview. Confirm the live support matrix before relying on macOS coverage. Local runtime protection remains preview on Windows.

Local AI discovery is now available for Windows

On eligible, onboarded Windows endpoints, Defender can surface supported local agents in Assets > AI agents, associate them with the device and user account, and expose supported remote and local MCP configuration such as server names, types, endpoints, and launch commands. Microsoft says the announcement expands coverage to more than 50 supported agents and MCP servers. Discovery is activity-based: an installed agent may not appear until Defender observes it running, so this inventory is not a complete software inventory.

For administrators, the immediate step is to confirm Defender for Endpoint Plan 2 licensing and endpoint onboarding, then review which agents and MCP servers appear in the inventory. The inventory and Advanced Hunting data are available with Plan 2; agent risk ratings, indicators, and recommendations require Microsoft 365 E7 or Microsoft Agent 365 together with Plan 2. Do not treat the discovery GA announcement as GA runtime blocking: local runtime protection remains preview, currently for supported Windows agents.

  • Verify Defender for Endpoint Plan 2 and active endpoint onboarding.
  • Review Assets > AI agents and inspect the associated user, device, and MCP configuration.
  • Compare your observed agents with Microsoft’s current support matrix; installation alone does not guarantee discovery.
  • Keep runtime protection in a scoped preview pilot and audit before considering block mode.

What changed for Microsoft customers

ChangeCurrent meaningAction
Agent 365 reached general availabilityMicrosoft now positions Agent 365 as the enterprise control plane for observing, governing, and securing agents across Microsoft and supported partner ecosystems.Decide which agents must be registered, who owns the registry, and how lifecycle status is enforced.
Jul 1 licensing transitionListed Copilot Studio and Foundry security capabilities require an Agent 365-eligible license.Verify entitlements, trials, Defender settings, and continuity of visibility.
Security for AI consolidatedMicrosoft documents a consolidated Security for AI Agents enablement path in Defender, supported by Agent 365 observability and registry data.Confirm connectors and supported agent platforms are actually onboarded.
Local agent discovery reached GA on WindowsMicrosoft announced Windows and macOS discovery for more than 50 agents and MCP servers. The current Learn matrix still labels macOS discovery as preview. Defender for Endpoint Plan 2 is the minimum for inventory and hunting; posture details have additional licensing requirements.Use the live support matrix to verify each platform and agent. Keep local runtime protection separate: it remains in preview.

The four layers Defender can provide

LayerWhat it helps answerTypical output
DiscoveryWhich supported agents exist, where they run, who or what identity they use, and which tools they can reach.AI Assets inventory, Agent 365 registry context, and Advanced Hunting data.
Posture and riskWhich agents combine risky access, weak instructions, autonomy, sensitive data, tool exposure, or active alerts.Risk levels, indicators, recommendations, relationships, and prioritization context.
Runtime protectionCan a supported risky prompt, tool call, or response be audited or blocked before the action completes?Audited or blocked behaviors, policy matches, and behavior records in Defender.
Detection and investigationWhat happened, which agent and user were involved, what tool was invoked, and what the possible blast radius is.Alerts, correlated incidents, entity relationships, Advanced Hunting queries, and response workflows.

How Microsoft assesses agent risk

Microsoft Defender’s agent posture model combines active risk indicators. These can come from an agent’s configuration, instructions, tools, permissions, runtime activity, endpoint and user context, or active security alerts.

Examples in Microsoft’s documentation include weak instructions, high usage, indirect prompt-injection exposure, privileged business-system access, and an active threat. For local agents, the surrounding context matters: running for a critical user, on a critical or vulnerable device, or with privileged software-development access can raise the potential impact.

A high-risk rating is not automatically proof that an agent was built incorrectly. A public support agent may intentionally accept unauthenticated traffic. An operations agent may need write access to a business system. Those facts still increase exposure, so the right question is whether the business purpose, permissions, monitoring, approvals, and compensating controls justify the risk.

Review risk level and recommendations together. Microsoft notes that a high-risk agent may have no available recommendation when an indicator is intentional or not directly remediable, while a lower-risk agent can still have a useful configuration recommendation.

Runtime protection has real coverage boundaries

Defender can inspect activity through the agentic loop and, for supported paths, block risky actions before execution. Coverage differs by agent type:

  • Agent 365 tool invocations: Microsoft documents protection through Work IQ MCP for supported tool calls, including supported customer MCP tools onboarded to Agent 365. Agents using unsupported tools or paths outside that integration are not covered automatically.
  • Copilot Studio agents: Defender evaluates supported tool invocations after Copilot Studio is connected in the Security for AI setup.
  • Local AI agents: These require separate endpoint onboarding. Defender for Endpoint must run in active mode for runtime protection on supported endpoints.

This is why an inventory cannot stop at the agent name. For each important agent, document its identity, data, tools, MCP servers, platform, runtime, user population, supported protection path, and any execution route Defender cannot see.

Start in audit before you block

Microsoft’s current design includes a default audit rule for cloud agents. Matching behavior is recorded without stopping the action, which gives the security team a baseline. Custom rules can then block selected detection types and be scoped to all or specific agents.

A controlled rollout should:

  1. Run in audit long enough to understand normal tool use and false positives.
  2. Choose high-confidence threats where blocking is safer than allowing execution.
  3. Test the block against a nonproduction or low-impact agent.
  4. Define who receives the event, who can disable the rule, and how business operations continue.
  5. Expand scope only after the agent owner and security team agree on the result.

Defender records audited and blocked activity as behaviors, including the agent, user, tool, and reason. That creates useful hunting and automation data, but it also changes alert behavior: Microsoft notes that near-real-time alerts continue in audit mode, while a blocking rule may prevent a separate near-real-time alert for the covered agent. Build dashboards and response playbooks with that distinction in mind.

Prompt evidence needs a privacy decision

Defender can include suspicious prompt or response snippets as alert evidence. Microsoft says sensitive data and secrets are redacted, but also cautions that the conversation itself may remain sensitive. Administrators can control prompt-evidence collection.

Before enabling broad collection, decide:

  • Which security roles can view prompt evidence
  • How incidents containing customer, employee, legal, health, or financial context are handled
  • What retention and export rules apply
  • How investigators avoid copying sensitive evidence into tickets and chat
  • Whether regional, labor, or contractual requirements change the design

A practical 30-day rollout

Week 1: Confirm the control plane

  • Validate Agent 365 licensing and the July 1 transition status.
  • Confirm Security for AI Agents is enabled as intended.
  • Connect the Microsoft 365 app path and Copilot Studio where applicable.
  • Identify local agents that require Defender for Endpoint onboarding.

Week 2: Reconcile the inventory

Compare the Agent 365 registry, Defender AI Assets, Copilot Studio environments, Foundry projects, endpoint discoveries, SaaS records, and procurement data. Assign an owner and business purpose to every material agent. An unowned agent should not retain broad access simply because it appears in a Microsoft inventory.

Week 3: Prioritize posture

Start with agents that can write to important systems, use maker or user credentials, access sensitive data, run without approval, serve many users, or already have active alerts. Reduce unnecessary permissions and tool scope before relying on detection to catch misuse.

Week 4: Tune runtime controls

Review audit behaviors, select a small set of high-confidence blocking scenarios, test response ownership, and document unsupported paths. Establish a monthly operating review for new agents, changed tools, risk movement, incidents, exceptions, and stale registrations.

What Defender does not replace

  • Identity design: Agents still need unique, scoped identities, least privilege, credential lifecycle, and Conditional Access where supported.
  • Data governance: Purview labels, DLP, retention, access reviews, and clean permissions remain necessary.
  • Secure build practices: Instructions, tool contracts, input validation, secrets handling, dependency security, and red-team testing remain the builder’s responsibility.
  • Business approvals: High-impact actions still need appropriate human review, separation of duties, and transaction controls.
  • Complete third-party coverage: Every agent platform, endpoint, MCP server, or tool path must be verified. Registration does not guarantee runtime inspection.

Frequently Asked Questions

What does Defender add to Agent 365? Defender provides the security operations layer: supported discovery, risk and posture, runtime audit or blocking, threat detection, incident correlation, and hunting.

Did licensing change? Yes. Microsoft states that the listed Copilot Studio and Foundry agent-security capabilities require an Agent 365-eligible license as of July 1, 2026.

Is local agent discovery generally available? Microsoft announced GA for discovery on Windows and macOS on Oct. 29, 2026, but its current Learn support matrix still labels macOS discovery as preview. The Learn matrix lists Windows agents in its standard Windows section and labels macOS discovery as preview; verify the live support matrix for each platform and agent. Local runtime protection remains preview.

What license is needed to inventory local agents? Microsoft Learn lists Defender for Endpoint Plan 2 as the minimum for local agent discovery and Advanced Hunting. Risk levels and recommendations require Microsoft 365 E7 or Microsoft Agent 365 together with Plan 2.

Does Defender protect every tool call? No. Coverage depends on the agent type, connection, supported tool path, and runtime. Unsupported tools or paths outside the supported integration are not automatically protected.

Should we enable blocking immediately? Usually not across the full estate. Begin with audit, understand normal behavior, and introduce tested blocking rules for high-confidence scenarios.

Related services

Want a second set of hands on this? Accred Consulting provides Microsoft 365 security consulting for US organizations.

Build an AI agent security operating model

Accred Consulting can help reconcile agent inventory, identity, permissions, Defender coverage, Purview controls, runtime policies, and incident ownership into one practical Microsoft 365 governance program.

Review Your Agent Security