Microsoft Agent 365 brings Shadow AI discovery into an agent governance workflow, helping teams identify unmanaged agents, assess their context and choose a response. It matters to security and IT teams that need to investigate AI tools running beyond their approved inventory without treating every finding as an automatic block. Microsoft Events demonstrates the workflow in an Ignite session, including agent discovery, risk review and control options.
The session frames agent security as a two-part task: protect agents the organization already knows about, then find activity beyond that inventory. For known agents, the demonstrated review connects identity and ownership with evidence such as data access, permissions and tool actions. The team can then select a supported response based on the agent, its role and policy.
Agent 365's registry is shown with a dedicated Shadow AI area. In the demonstration, it surfaces detected tools and gives the administrator a way to investigate individual instances. An agent identifier can help distinguish a tool across devices; the instance view adds details such as its device, operating system, model, tools and dynamic risk assessment.
That extra context matters because the same agent name can represent different activity on different machines. In the example, a Goose Desktop instance is associated with a laptop and tools including DeepWiki, Git and GitHub. The presenter highlights the possibility that a local tool could send information to a remote MCP server, making tool connections part of the risk review rather than treating the installed agent as the whole picture.
The workflow is aimed at security and IT teams responsible for agent visibility, investigation and policy response. It is also relevant to administrators who manage devices and need to consider whether an agent should be blocked, restricted or investigated. The demonstration begins in the Agent 365 registry and moves from an agent listing to an individual instance.
The demonstrated controls vary by agent and environment. Blocking is described as using policies pushed through Intune to prevent an agent or its binaries from running on managed systems. The session also shows restrict options marked Preview: changing device status, restricting network traffic and isolating an agent. Treat those options according to their displayed status and the environment in which you manage them.
Partner demonstrations add other approaches, but they are not the same as the Agent 365 registry controls. Zenity shows intent-based session monitoring and enforcement for GitHub Copilot, while Check Point shows its Agent Behavior Defense approach. These examples illustrate partner capabilities presented alongside Microsoft's; they should not be read as identical built-in Agent 365 functions.
A practical triage sequence from the demonstration is to identify the agent, examine evidence and context, then choose an eligible protection. The registry helps start with known agents and expand the review to Shadow AI. Once a finding appears, the instance details help the team decide whether to investigate, apply a control or take another action.
The example also shows activity information, including instance and user activity over the previous 30 days. If Global Secure Access is in use, the presenter says the view can include sites visited by the agent. The session further describes proactively applying policy to protect against an agent instance that has not yet been detected.
Not every response is suitable for every agent or environment: the presentation explicitly says available controls vary. In particular, the options to change device status, restrict network traffic and isolate an agent are labeled Preview in the demonstration. Blocking is a more direct response, while restriction can target device compliance or network destinations; the choice depends on the risk and the control available to that agent.
Discovery and inventory do not, by themselves, explain every action an agent takes at runtime. The partner examples address that separate concern. Zenity classifies a session as production operations and demonstrates a rule that blocks an attempt to handle a secret in that context. Check Point demonstrates judging an action against the agent's task and blocking or steering behavior. Those examples underline a useful distinction: knowing which agent is running is different from evaluating what it is trying to do.
Local tools and remote MCP servers deserve attention during investigation because they can extend an agent's reach beyond its basic installation. Review the tools, data access and device context before deciding whether to approve, restrict or block an instance. The session presents governance as a continuing cycle of discovery, assessment and response, rather than a one-time approved-tools list.
The demonstration shows agent and instance identifiers, risk information, and context such as device, operating system, model and tools. The detail available can vary by agent and environment.
The session demonstrates blocking through Intune policies that prevent agents or their binaries from running on systems. It also shows preview options for restricting or isolating an agent.
Changing device status, restricting network traffic and isolating an agent are labeled Preview in the demonstration.
The demonstration shows instance-level device, model, tool and risk context, plus an option to create an investigation case. ServiceNow and Azure DevOps integrations are named for case creation.