Disable Copilot actions and agents in Windows and Microsoft 365
"Copilot actions" now means several different things: an agent that clicks through Windows apps, agents that work inside Outlook and Teams, a browser mode that acts on web pages and a tool that drives a computer for Copilot Studio agents. Each has its own off switch, in a different console.
Here is the map, the switches and what to do about the agents Microsoft's settings never see.
AI Agent Blocklist: cover the agents your Microsoft switches do not see
AI browsers, agent extensions, desktop agents and coding agents from every vendor, in files for Intune, Defender, Group Policy, Edge and your firewall.
From $99 a month for one organisation
Which Copilot actions are you dealing with?
Start by naming the feature. People search for "disable Copilot actions" when they mean very different things, and the wrong switch leaves the right agent running.
| Feature | Where it runs | What it can do | Main off switch |
|---|---|---|---|
| Copilot Actions on Windows | Copilot app, Copilot Labs, Windows Insider builds | Clicks, types and scrolls in desktop apps; works with files in known folders | Settings > System > AI Components > Experimental agentic features (off) |
| Agent connectors on Windows | On-device registry of MCP servers | Gives agents access to apps and system tools | WindowsAI policy ConfigureAgentConnectors = 2 |
| Agents in Microsoft 365 Copilot | Copilot, Teams, Outlook, Microsoft 365 apps, SharePoint | Search, act through connectors, APIs and plugins | Microsoft 365 admin center > Agents > Settings > User access |
| Computer use in Copilot Studio | A Windows machine set up for the agent | Operates websites and desktop apps with a virtual mouse and keyboard | Access control allow list per tool, plus who may build agents |
| Copilot in Edge | Microsoft Edge | Chat in the side pane; Copilot Mode can act on web pages | Edge policies, see our Edge guide |
| GitHub Copilot coding agent | GitHub Actions environment or cloud sandbox | Writes code and opens pull requests | GitHub organization and enterprise policies |
Other vendors have used the same words for their own assistants. Those are managed in each vendor's own admin console, not in Windows or Microsoft 365.
The problem: agents that act with a person's rights
A chat assistant suggests. An agent does. Once Copilot or any other agent can click, send and change things, mistakes and manipulated instructions turn into real changes in real systems.
Hidden instructions
Microsoft names cross-prompt injection as a risk: text hidden in a document or a page that overrides what the user asked the agent to do.
Data leaving unnoticed
An agent that can read files and send messages can move data out in ways no one would approve as a single request.
Unwanted installs
Microsoft also lists malware installation among the unintended actions a manipulated agent could take.
Agent sprawl
Staff can build and share agents with Agent Builder. Owners leave, agents stay, and nobody reviews what they connect to.
Shared credentials
A Copilot Studio computer use tool with maker credentials lets everyone who uses the agent act with the maker's access on that machine.
Agents outside the tenant
Third-party AI browsers and agent extensions never appear in Microsoft's agent lists, so a clean Registry can hide a busy laptop.
Signs you need to act now
Copilot Actions on Windows: how it works and how to keep it off
Copilot Actions is an experimental feature for Windows Insiders, started from Copilot Labs in the Copilot app. It uses vision and reasoning to work through apps for the user. Windows runs it apart from the user's own session.
The user's session
- The person's own account and desktop
- Their apps, browser profiles and signed-in services
- Their full set of files
The agent workspace
- A separate agent account per agent, so its actions can be told apart
- A separate, lighter session than a full virtual machine
- Read and write access only to Documents, Downloads, Desktop, Music, Pictures and Videos, when granted
- The gate is Settings > System > AI Components > Experimental agentic features. It is off by default and creates the agent account and workspace only when switched on.
- Only an administrator of the device can turn it on, and the choice then applies to every user on that device. Removing local admin rights from staff keeps it off.
- On recent preview builds each agent has file settings under System > AI Components > Agents, with Allow Always, Ask every time and Never allow.
- Keep Insider builds off production laptops. The feature ships through the Insider program first, so your update rings are a control too.
Check the current policy list: Microsoft's support article for this setting names no Group Policy or Intune setting for the toggle itself at the time of writing. Admin rights and update rings are the dependable controls today.
WindowsAI policies that matter for agents
These settings are in the WindowsAI area of the Policy CSP and in Group Policy under Windows Components. Several apply only to Insider Preview builds on Enterprise and Education editions, and Microsoft marks them as subject to change.
| Policy | What it does | Values | Applies to |
|---|---|---|---|
| ConfigureAgentConnectors | Turns agent connectors (MCP servers for agents) on or off | 0 user in control, 1 force enable, 2 force disable | Insider Preview, Enterprise and Education |
| AgentConnectorAccessPolicy | Lists which MCP server and host connections are allowed, as a JSON string | Your allow list | Insider Preview, Enterprise and Education |
| AgentConsentDuration | How long an agent consent lasts before Windows asks again | 1 to 8760 hours, default 720 | Insider Preview, Enterprise and Education |
| DisableSettingsAgent | Turns off the agentic natural language search in the Settings app | 1 disables it | Insider Preview, Enterprise and Education |
| OnDeviceRegistryLoggingLevel | Sets logging detail for the on-device agent registry | 0 errors, 1 information, 2 debug | Insider Preview, Enterprise and Education |
| RemoveMicrosoftCopilotApp | Uninstalls the Microsoft Copilot app when Microsoft 365 Copilot is also installed, the user did not install it and it was not opened in 28 days | 1 removal enabled | Windows 11 24H2 and later |
| TurnOffWindowsCopilot | The old switch for Copilot in Windows. Deprecated, and it does not cover the new Copilot app | 1 disables | Older builds |
Tip: set ConfigureAgentConnectors to 2 on test rings now, so connectors stay off when agent features reach your normal builds.
Pair it with OnDeviceRegistryLoggingLevel at 1 on pilot devices to see which connectors agents try to use. Our guide to disabling MCP servers covers the same idea for other agent apps.
Agents in Microsoft 365: the admin center switches
Agent management is on by default in tenants licensed for Microsoft Copilot. Everything below lives on the Agents pages of the Microsoft 365 admin center, and changes sync with the Teams admin center when unified app management is on.
- See what exists. Go to Agents > All agents > Registry. The summary shows Total agents, Agents without owners and Unmanaged agents. Use Export for a CSV you can review with owners.
- Filter by channel. The Channel filter shows where each agent is deployed: Copilot, Teams, Outlook, Microsoft 365 apps or SharePoint. Start with agents that reach mail and chat.
- Limit who can use agents. Under Agents > Settings > User access, the default is All users. Choose No users to stop agents, or Specific users or groups for a pilot.
- Limit which agents exist. In Agent and plugin access, switch off agents and plugins built by external publishers and, if needed, those built by your organization. Microsoft-built agents stay visible but cannot be installed when that option is off.
- Stop company-wide sharing. Sharing offers All users, No users or Specific users, and applies to agents built with Agent Builder.
- Clean up in bulk. Agent management rules include a rule to block ownerless agents without usage and one to reject old publish requests.
- Review tools and MCP servers. Plugins, skills and MCP servers are managed under Agents > Tools, separate from the agent list.
Not covered: Microsoft states that Researcher and Analyst are part of the core Copilot chat experience and do not fall under the agent settings above. Decide on them through your Copilot licensing and policies instead.
- User access: No users
- External publisher agents: off
- Organization-built agents: off
- Sharing: No users
- Block every agent in the Registry you did not publish
- User access: one pilot group
- External publisher agents: off
- Organization-built agents: on, with admin approval
- Sharing: Specific users
- Monthly rule to block ownerless agents without usage
For listing agents in a script, Microsoft Graph offers a preview catalog API for Copilot packages that works with the AI Administrator role and the CopilotPackages.Read.All permission. Our page on building an AI agent inventory shows how to combine that list with device and network data.
Computer use in Copilot Studio: the settings to set before anyone publishes
Copilot Studio's computer use tool lets an agent operate websites and Windows desktop apps with a virtual mouse and keyboard. By default it can operate on any website or application on the machine it uses, so the per-tool settings matter.
Turn it on and list the exact websites and desktop apps the tool may act on. Other sites can still open, but actions on them fail.
Keeps the tool off plain HTTP sites so typed data does not travel unencrypted.
Maker-provided credentials are the default. Share such an agent and every user acts with the maker's access on that machine.
Name a reviewer who gets an email when the tool detects possibly harmful instructions, with a time limit to respond.
Machines
Use dedicated machines with a least-privilege account, a browser allow list and application control. Microsoft's own guidance says the same.
Stored passwords
Credentials can sit in Power Platform storage or in an Azure Key Vault you own. Prefer the vault, and scope each credential to one login domain.
The wider risks of screen-driving agents, from any vendor, are on our page about computer use agent security.
Copilot in Edge and Copilot on GitHub
Edge
The policy Microsoft365CopilotChatIconEnabled (Edge 139 and later) controls whether the Microsoft 365 Copilot Chat icon shows in the Edge for Business toolbar for Entra ID profiles.
Copilot Mode, the agentic browsing mode, has its own controls. Our guide to disabling Copilot Mode in Edge walks through them.
GitHub
GitHub Copilot's coding agent runs sessions in a GitHub Actions based environment or a cloud sandbox and opens pull requests. Organization and enterprise policies decide whether it is available, and repository admins can turn Copilot off per repository.
How we solve the part Microsoft's switches cannot reach
Windows and Microsoft 365 settings govern Microsoft's agents and the agents in your tenant. Staff can still install an AI browser, add an agent extension or run a desktop agent from another vendor. The AI Agent Blocklist closes that gap.
Covered by Microsoft settings
- Copilot Actions and agent connectors on Windows
- Agents in Copilot, Teams, Outlook and SharePoint
- Copilot Studio agents and their computer use tools
- Copilot Chat in Edge for Business
Covered by the AI Agent Blocklist
- AI browsers and browser agent extensions from any vendor
- Desktop agents and personal assistants on Windows and macOS
- Coding agents, their CLI packages and MCP servers
- Agent platforms and agent builders on the web, by domain
Microsoft's own consumer Copilot is in the files too, for teams that standardise on Microsoft 365 Copilot and want the consumer app gone. Examples: the Windows package Microsoft.Copilot and the domain copilot.microsoft.com.
Deploy it next to your Microsoft settings
- Import our extension block lists into Intune Settings catalog policies: Control which extensions cannot be installed for Edge and Configure extension installation blocklist for Chrome.
- Load our domain file into Microsoft Defender as custom indicators, or into your firewall, to block agent sites and APIs in every app. See blocking AI sites with Defender.
- Use our Windows package IDs and process names to find agent apps in the Intune Discovered apps report and to build App Control rules. Full steps in block AI with Intune.
- Pick the risk grades you block. Some teams block every agent; others block higher-risk entries and allow approved tools, as explained in allowlist or blocklist.
Block
- Experimental agentic features on production devices
- External publisher agents in Microsoft 365
- Third-party AI browsers and agent extensions
Monitor
- Agents without owners and Unmanaged agents counts
- New agent connectors on pilot devices
- New entries in our daily list updates
Allow
- Microsoft 365 Copilot for licensed staff
- Agents your organization published after review
- Computer use tools with an allow list and a vault
Meeting bots are a separate case with their own settings: see blocking AI bots in Teams meetings.
Every agent on one list, updated daily
Microsoft's controls handle Microsoft. Our files handle the rest of the agent market, which changes every week. New AI browsers, extensions, desktop agents and MCP servers are added as they appear.
See what each file contains on the AI Agent Blocklist page, then pick a plan.
See plans and prices- 2,600+ AI agent domains across agent categories
- 5,300+ device identifiers: extension IDs, Windows package IDs, macOS bundle IDs, process names, CLI packages
- Name, category, identifier type and risk grade on every entry
- Intune, Group Policy and Google Admin ready extension lists
- Plain, CSV, JSON, hosts, PAC and firewall list formats
- Access set up within one business day after your order
Ask us or order a plan
Tell us what you need to block and where your rules live. We reply within one business day.
- Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
- Updated every day as new AI browsers, agents and extensions appear
- Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you
Copilot actions FAQ
What is Copilot Actions?
On Windows, Copilot Actions is an experimental agent in Copilot Labs for Windows Insiders. It clicks, types and scrolls in apps for the user, inside a separate agent workspace under its own agent account.
The phrase is also used loosely for agents that act in Microsoft 365, for Copilot Mode in Edge and for other vendors' assistants.
How do I disable Copilot Actions on Windows?
Keep Settings > System > AI Components > Experimental agentic features turned off. It is off by default and only an administrator can turn it on, so standard users without admin rights cannot enable it.
On Insider builds you can also force-disable agent connectors with the WindowsAI ConfigureAgentConnectors policy set to 2.
How do I enable Copilot Actions or allow actions for a pilot?
An administrator turns on Experimental agentic features on the test device, which applies to every user of that device. Then the user starts Copilot Actions from Copilot Labs in the Copilot app. Approve folder access per agent and keep the pilot on Insider test machines with no production data.
How do I stop staff using agents in Microsoft 365 Copilot?
In the Microsoft 365 admin center go to Agents > Settings > User access and choose No users or Specific users or groups. Use Agent and plugin access to switch off agents built by your organization or by external publishers, and block individual agents in the Registry.
Researcher and Analyst are part of core Copilot chat and are not covered by agent settings.
What are Copilot actions in Outlook?
Agents can be deployed into Outlook as a channel, alongside Copilot, Teams, Microsoft 365 apps and SharePoint. In the agent Registry, filter by the Outlook channel to see which agents reach mailboxes, then block or remove the ones you have not approved.
What does copilot actions github mean?
GitHub Copilot's coding agent runs its sessions in a GitHub Actions based environment or in a cloud sandbox. It is controlled by GitHub organization and enterprise policies and per repository, not by Windows or Microsoft 365 settings.
Can I see which actions Copilot took?
On Windows, agents run under their own agent account so their actions are distinguishable from the user's. In Microsoft 365, the agent Registry export and usage export show which agents exist and who uses them.
For detailed records of agent activity, check your audit and compliance tooling in the current admin portals.
Does turning off Copilot agents stop other AI agents?
No. Microsoft's switches govern Microsoft's agents and the agents in your tenant. AI browsers, browser agent extensions, desktop agents and coding agents from other vendors need device, browser and network blocks, which is what the AI Agent Blocklist provides.
Start with our list of places shadow AI agents hide.
Microsoft, Windows, Microsoft 365, Copilot, Copilot Studio, Outlook, Teams, SharePoint, Edge, Intune, Defender, Entra, Power Platform and Azure are trademarks of the Microsoft group of companies.
GitHub and GitHub Copilot are trademarks of GitHub, Inc. Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.