Microsoft Defender for Endpoint

Defender block AI sites: stop AI agents and AI browsers before they connect

AI agents reach mail, files and business apps through domains your web rules have never heard of. New ones appear every week.

Defender for Endpoint can block them on every managed device, on or off your network, once it has the list.

Indicators, URLs/Domains
Example: AI agent domains by action
manus.imAutonomous agentBlock
browser-use.comBrowser agentBlock
genspark.aiAutonomous agentWarn
smithery.aiMCP serversAudit
your approved AI toolLicensedAllow
Next step

AI Agent Blocklist: the domains and app identifiers Defender needs

AI agent domains to import as custom indicators, graded by risk so they fit your indicator budget, plus Windows app identifiers for app rules. Updated every day.

From $99 a month for one organisation

Why AI agents get past the web rules you already have

Most Defender setups were built for malware and for the browsers people used five years ago. AI agents arrive as new browsers, desktop apps and command-line tools, and each one is enforced by a different part of Defender.

Where the AI traffic comes fromBrowser URL policyWeb content filteringCustom indicators with network protectionApp control
Microsoft EdgeYesYes, through SmartScreenYes, through SmartScreenNot needed
Chrome, Firefox, Brave, OperaOnly if managedYes, through network protectionYesNot needed
AI browsers (own process name)NoNot a supported browserYes, as a non-browser processYes, stops it running
Desktop agent appsNoNoYesYes
Command-line coding agentsNoNoYesPartly, scripts run under node or python

Microsoft documents that web content filtering identifies supported browsers by process name: Edge, Chrome, Firefox, Brave and Opera. An AI browser runs under its own name, so domain indicators and app rules are the controls that reach it.

Unknown Chromium browsers in your software inventory next to Edge and Chrome

Network events from agent apps to domains no one approved

node or python processes calling AI agent services on staff laptops

New sign-ins to AI tools with company accounts, seen in your identity logs

How Defender custom indicators block AI domains

A custom indicator tells every onboarded device what to do with a domain, IP or URL. For AI agent sites, domain indicators are the right tool: one entry covers the domain and all its subdomains.

Prerequisites to check first

  • Microsoft Defender Antivirus in active mode, with cloud-based protection and behavior monitoring on
  • Network protection in block mode for browsers other than Edge and for non-browser processes
  • "Custom network indicators" turned on in Settings > Endpoints > General > Advanced features
  • SmartScreen on in Edge, which enforces indicators there
  • Devices on a current antimalware platform version; check the current admin console for the minimum
Domain covers subdomainsAn indicator for a domain matches it and every subdomain under it.
Allow beats Warn beats BlockIf one domain has several indicators, Allow wins, then Warn, then Block.
Full HTTPS paths: Edge onlyOther browsers are matched on the domain name, so path rules are not portable.
Up to 48 hoursMicrosoft says most indicators apply in under two hours; allow two days.
Many platformsDomain indicators are supported on Windows, macOS, Linux, iOS and Android.
Scoped by device groupEach indicator applies to all devices or to the device groups you pick.

Network protection reads the domain name from the TLS handshake. Traffic over QUIC or with Encrypted Client Hello hides it. In Chrome, the QuicAllowed and EncryptedClientHelloEnabled policies turn these off. Many teams also block outbound UDP 443 at the firewall.

On Macs and Linux devices, network protection enforces indicators and web content filtering in every supported browser, Edge included. Turn it on in block mode there too, or Mac users with an AI browser will not see your blocks.

Plan around the 15,000 indicator limit

Microsoft documents a hard limit of 15,000 indicators per tenant, shared by files, certificates, IPs and URLs. It cannot be raised. AI domains have to fit next to everything your security team already tracks.

15,000indicators per tenant, all types together
2,600+AI agent domains in our agent list
12,000+AI tool domains in our wider AI data, too many to add whole
AI agent domains, about 2,600 Your threat intelligence (example) File and certificate indicators (example) Room left for new entries

How to stay inside the budget

  • Add the AI agent list, not every AI tool domain on the web
  • Filter by risk grade: Severe and High first, Elevated next
  • Use one domain indicator per tool, never one per subdomain
  • Set expiry dates on test indicators so they leave by themselves

When the list is bigger than the budget

  • Put the full AI list on your DNS filter, where size is not limited this way
  • Load it on the firewall as an external list
  • Keep Defender for the agent domains that must follow laptops off the network
  • Sync changes from our daily update rather than reloading everything

Step by step: load the AI Agent Blocklist into Defender

The steps below use the Microsoft Defender portal. Menu names change from time to time, so check the current admin console if a label differs.

  1. Turn on network protection in block mode

    Most teams set this in Intune under Endpoint security > Antivirus. Our guide to blocking AI with Intune covers the profile.

  2. Turn on Custom network indicators

    In the Defender portal open Settings > Endpoints > General > Advanced features and switch it on.

  3. Pick the slice of the list

    From our CSV, choose agent domains by category and risk grade. Keep tools your organisation has approved out of the block set.

  4. Fill the import template

    Open Settings > Endpoints > Indicators, tab IP addresses or URLs/Domains, choose Import and download the current CSV template from the import panel. Map our domain, tool name and category into its columns.

  5. Scope and import

    Start with a pilot device group. Give every indicator a clear title, such as the tool name, so alerts make sense to the analyst who reads them.

  6. Test on a pilot device

    Wait for the policy, then open a blocked domain in Edge, in Chrome and from an AI browser. Edge shows a block page; other processes get a system notification.

  7. Keep it current

    Our list changes every day. Apply the additions and removals on a schedule, by import or through the Defender API, and widen the scope once the pilot is clean.

Warn, audit or block: one action per risk grade

URL and domain indicators support four actions: Allow, Audit, Warn and Block execution. The risk grade on every entry in our list maps cleanly to them.

Grade in our listSuggested actionWhat the user seesWhy
SevereBlock executionBlock page in Edge, notification elsewhereThe highest grade in the list: block these first
HighBlock executionSame as aboveBlock next, before staff find and adopt them
ElevatedWarn, with a redirect URL to your AI policyA warning they can bypassSteers people to approved tools without stopping work
ModerateAuditNothingBuilds a picture of real use before you decide
ApprovedAllowNormal accessOverrides blocks, because Allow wins

For Warn to work in browsers other than Edge, users must see the "Files or activities are blocked" notifications in Windows Security. The warning sets a bypass period, and you can point the redirect URL at your AI acceptable use page.

Block the AI browser itself, not only its sites

An AI browser that is installed keeps working on every site you did not block. Stopping the app is the cleaner fix. Defender gives you three ways, and Microsoft recommends one of them for apps.

Not recommended

File hash indicators

Each version of an app has a new hash, and apps are made of many files. Microsoft advises against using hashes to block applications.

Useful only as a short stop-gap for one known file.

Use with care

Certificate indicators

Allow or block and remediate files signed with a given certificate, supplied as a .CER or .PEM file.

It hits everything that publisher signs, so check what else uses that certificate first.

Recommended

App Control for Business or AppLocker

Microsoft's preferred way to block applications. Deny rules or an allow list stop AI browsers and agent apps from starting.

Deploy through Intune and start in audit mode.

Identifiers from our list you can build rules and hunting queries on

  • Comet, Windows package ID Perplexity.Comet
  • Cursor, Windows package ID Anysphere.Cursor
  • Codex CLI, process name codex-x86_64-pc-windows-msvc.exe
  • BrowserOS, process name browseros-claw-server.exe

These are a few examples. The download has the full set of AI browser and agent app identifiers for Windows and macOS. In advanced hunting, query process and network events for them to see who runs what.

A blocked connection can still appear as ConnectionSuccess in network events, because network protection decides after the TCP handshake. Check the matching alert to confirm the block. For removal of installed AI browsers, see how to block AI browsers.

Cursor blocked by Windows Defender? Find out which kind of block it is

Developers often report that Defender stopped Cursor or its updater. Before you change anything, find out whether it is a reputation warning, a detection or a rule your own team set. Each needs a different answer.

A SmartScreen warning on a new installer

New installer releases start with little download reputation, so SmartScreen can show an unrecognised app warning even when the file is signed. This fades as the release gains reputation.

In a company, ship approved versions through your software deployment instead of letting each user download them.

A detection on a new or temporary file

Users have reported detections on installer temporary files and on new CLI builds that the vendor called false positives. Read the alert and protection history before acting.

Do not answer with a broad folder exclusion in the user profile. Submit the file to Microsoft for review, or add a narrow allow indicator for the verified file.

A rule your own security team set

An indicator for cursor.com, an App Control rule or an application block set in Defender can be deliberate. Check your own indicators and policies before you raise a ticket with anyone else.

If the block is policy, the developer needs an approved exception, not a workaround.

Deciding whether to allow it

Cursor is a coding agent that reads source code and can run commands. Decide per team, write the decision down, and enforce it the same way on every device.

Our page on blocking or allowing Cursor lists the controls for both choices.

What Defender indicators cannot block

Defender is strong on devices you manage. Know its edges so you can cover them with something else.

  • Microsoft-signed apps: file and certificate indicators cannot block them, so Copilot needs Microsoft's own policies
  • QUIC and Encrypted Client Hello traffic in browsers where you have not turned them off
  • Full HTTPS URL paths outside Edge, and sites loaded through HTTP/2 connection reuse
  • AI features inside domains you allow, such as an assistant built into a business app
  • Devices that are not onboarded, including personal phones and contractor laptops
  • Internal IP addresses, which cannot be indicators at all

For Microsoft's own AI, use the Edge Copilot Mode settings and the Copilot Actions controls. For agents that connect tools through MCP, see blocking MCP servers.

A four-week rollout that does not break work

Blocking hundreds of domains on a Monday morning creates a flood of tickets. Stage it, and let the data from audit mode tell you where the exceptions are.

Week 1Audit

Import the agent domains as Audit for all devices. Watch which tools appear, on which device groups, and how often.

Week 2Decide

Agree the approved AI tools with the business. Add Allow indicators for them and write the exceptions down.

Week 3Warn

Switch Elevated tools to Warn with a redirect to your AI policy. Block Severe and High tools for a pilot group.

Week 4Block

Extend blocks to all devices, add app control for AI browsers, and put the daily list update on a schedule.

The audit week doubles as a shadow AI review. Our guide to shadow AI agents covers what to do with what you find.

The product

Keep Defender's AI rules current without the research

The AI Agent Blocklist tracks AI browsers, agents, extensions and MCP servers as they appear, and gives you each one as a domain and as device identifiers.

From $99 a month for one organisation. After an order we set up your download access within one business day.

CSV and JSON with tool, category and risk grade
Plain domain list for indicator imports
Windows package IDs and process names
macOS bundle IDs
Chrome and Edge extension IDs
Firewall, DNS and PAC formats

Want to see which tools are covered first? Browse the list of AI browsers we track.

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

We use your details only to answer you. No newsletters.

Defender and AI sites: common questions

Can Microsoft Defender block AI sites?

Yes. Defender for Endpoint custom indicators can block, warn on or audit domains and URLs. In Edge they are enforced through SmartScreen; in other browsers and non-browser processes, by network protection.

Network protection must be on in block mode, and Custom network indicators must be turned on in Advanced features.

Does a domain indicator also block subdomains?

Yes. An indicator for a domain also matches its subdomains at every level. Full HTTPS URL paths are only matched in Microsoft Edge, so use domains for AI agent sites.

How many indicators can I add in Defender for Endpoint?

Microsoft documents a limit of 15,000 indicators per tenant, and it cannot be raised. File, certificate, IP and URL indicators share it, so choose AI domains by risk grade.

Why does web content filtering not stop an AI browser?

Web content filtering identifies supported browsers by process name: Edge, Chrome, Firefox, Brave and Opera. An AI browser runs under its own name.

Custom indicators enforced by network protection also cover non-browser processes, and app control rules can stop the AI browser from running at all.

Why is Cursor blocked by Windows Defender?

Common causes are a SmartScreen warning on a new installer, a detection on a new or temporary file that the vendor treats as a false positive, or a rule your own team set.

Check the alert or protection history first, then follow your organisation's policy for coding agents.

How long do Defender indicators take to apply?

Microsoft says a new URL or IP indicator can take up to 48 hours to reach devices, and in most cases under two hours. Web content filtering policies can take up to two hours.

Does Defender block AI sites on Macs?

Yes, if Defender for Endpoint is onboarded on the Mac and network protection is on in block mode. On macOS, network protection enforces domain indicators in supported browsers, including Edge.

Add app rules or removal scripts for AI browsers installed on Macs, since the app itself keeps running until you remove it.

Can Defender block Microsoft Copilot?

Microsoft documents that Microsoft apps cannot be blocked by Defender file or certificate indicators, because they are signed by Microsoft. Use Microsoft's own Copilot and Edge policies for Copilot features.

Microsoft, Microsoft Defender, Windows, Intune, Microsoft Edge, SmartScreen, AppLocker and Copilot are trademarks of Microsoft Corporation. Google and Chrome are trademarks of Google LLC. Firefox is a trademark of the Mozilla Foundation.

Perplexity and Comet are trademarks of Perplexity AI, Inc. Codex is a trademark of OpenAI. Cursor is a trademark of Anysphere, Inc. Brave, Opera, Manus, Genspark, Browser Use, Smithery and BrowserOS are trademarks of their respective owners.

Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.