Firewall guide
Block AI on your firewall: AI agents, AI browsers and the services they call
AI browsers, agents and MCP servers send company data to services nobody approved, and they act with your users' logins. Your firewall already sees that traffic. It only needs a list of what to stop.
Where an agent request can be stopped
Next step
AI Agent Blocklist for firewalls
2,600+ AI agent domains in the external list format your firewall reads, updated every day.
From $99 a month for one organisation
What AI agent traffic looks like on your network
"AI traffic" is not one thing. Chat assistants are only part of it. The tools that worry security teams most are the ones that act: they open pages, fill forms, read mail and run commands. Each type leaves a different trail at the firewall.
AI browsers
Comet, ChatGPT Atlas, Dia, Opera Neon and similar browsers send page content and tasks to their own backends.
Seen as: new browser plus steady calls to the vendor's agent servicesAgent extensions
Extensions that click, type and read pages inside Chrome or Edge, using the user's open sessions.
Seen as: calls from a normal browser to an extension's APICoding agents
Claude Code, Codex CLI, Cursor and other agents that read code and run shell commands on developer machines.
Seen as: model API traffic from terminals and IDEsRemote MCP servers
Hosted servers that give an agent tools: mail, files, CRM, tickets, web scraping, payments.
Seen as: HTTPS sessions to MCP endpoints and directoriesAutonomous agents
Web agents such as Manus or Genspark, meeting bots and sales agents that work in the cloud for a user.
Seen as: logins and uploads from users to agent platformsWhat a firewall can see, and what it cannot
A firewall blocks by what it can read. For AI agents, the domain is almost always enough. The table shows where a domain list works, where TLS inspection adds detail, and where you need a device policy instead.
| What you want to stop | Without TLS inspection | With TLS inspection | Better control |
|---|---|---|---|
| AI browser reaching its agent backend | Yes, by DNS name and server name | Yes, plus full URLs in logs | Firewall list, and remove the app with MDM |
| Agent extension calling its API | Yes, by domain | Yes | Firewall list plus extension block list |
| Hosted MCP server used by an agent | Yes, by domain | Yes, by path as well | Firewall list plus MCP settings in each tool |
| A feature inside a site you allow | No, same domain | Often, by URL path | Browser or app policy for that feature |
| Local MCP server on the laptop | No, never leaves the device | No | MCP allowlist in each client, device control |
| Which extension is installed | No | No | Group Policy, Intune or Google Admin console |
| Agent traffic on a home network | No, unless tunnelled back | No | Always-on VPN, cloud gateway or DNS agent |
Signs AI agents are already on your network
Most teams find agents before they write a policy. These are the patterns admins tell us they spot first in firewall, DNS and proxy logs.
- DNS queries for AI browser update and sync hosts from machines that should run only your standard browser
- Long-lived HTTPS sessions to agent platforms during the working day, often from finance, sales and support
- Traffic to MCP directories and hosted MCP endpoints from developer subnets
- Uploads to deep-research or agent platforms right after users open internal file shares
- Model API calls from build servers or jump hosts where no one planned a coding agent
- Meeting bot domains joining calls that carry customer or patient data
- New hostnames with "agent", "mcp", "browser" or "operator" in them that no category in your rules matches yet
- Help desk tickets asking why a new browser cannot sign in to company apps
Who is exposed when an agent gets through
An agent acts with the rights of the person who started it. So the risk follows the user's access, not the tool's popularity.
Finance and purchasing
Agents that can open invoices, banking portals and supplier sites can be steered by text hidden in a page.
Executives and assistants
Mail and calendar access is what personal AI assistants ask for first, and that inbox holds the most sensitive threads.
Developers
Coding agents read source code and secrets, run commands and connect to MCP servers with production tokens.
Support and sales
CRM and ticket data flows into agents that summarise, reply and update records without a second look.
Regulated teams
Health, legal and HR data sent to an unapproved agent service is a data transfer you cannot document.
Schools and districts
Students use AI browsers and homework agents on school networks, often signed in with school accounts.
How we solve it: one list your firewall can read
The AI Agent Blocklist is a focused list of AI agent services, not a general AI category. It covers the tools that act on their own. Every entry is checked before it enters the list, and the lists update every day.
- AI browsers and browser agents
- General autonomous agents
- Personal AI assistants
- Agent-building platforms
- Multi-agent orchestration
- MCP servers and integrations
- Autonomous coding agents
- Deep-research agents
- Meeting and voice agents
- AI sales agents
- AI RPA
For the firewall
- 2,600+ AI agent domains in a plain list and in the external list format firewalls read from a URL
- CSV and JSON with tool name, category and risk grade (Low, Moderate, Elevated, High, Severe)
- DNS filter import files, a hosts file and a PAC file for proxies
- A risk grade on every entry, so you can start with the highest grades
For what the firewall cannot see
- 5,300+ agent identifiers on devices: extension IDs, macOS bundle IDs, Windows package IDs and process names
- Chrome and Edge extension block lists for Group Policy, Intune and the Google Admin console
- Command-line coding agents with their package names, such as @openai/codex
- App identifiers such as ai.perplexity.comet for Jamf and Intune
Pick your firewall: where the list goes
Each firewall has its own name for a list it can import or pull. The step-by-step guides show the menu paths, the rule order and how to test. Menus move between releases, so check the current admin console if a label differs.
Close the side doors before you trust the block
A domain block only works if traffic reaches the firewall in a form it can read. Check these four points on a test machine first. They are standard network hygiene, not special AI settings.
Encrypted DNS in the browser
If browsers use their own DNS over HTTPS, your DNS filter never sees the lookup. Chrome and Edge have the DnsOverHttpsMode policy to control this.
QUIC on UDP 443
Many firewalls inspect QUIC less fully than TCP. Blocking UDP 443 makes browsers fall back to TCP. Chrome also has the QuicAllowed policy.
Encrypted Client Hello
ECH can hide the server name in the TLS handshake. Where you rely on the server name, Chrome's EncryptedClientHelloEnabled policy can turn it off for managed browsers.
Traffic that leaves the office
Laptops at home skip the office firewall. Use an always-on VPN, a cloud gateway or a DNS agent so the same list follows the device.
A rollout plan that does not flood the help desk
Block in stages. Each stage gives you log data for the next one, and nobody loses a tool they rely on without warning.
- Days 1 to 7
Log only
Load the list with a monitor or log action. Export hits per user group and per category. You now know which agents are in use and by whom.
- Days 8 to 14
Block the highest risk grades
Block the entries graded Severe and High, filtered from the CSV. These are the agents with the most access and the least control. Tell users what changed and why.
- Days 15 to 21
Block AI browsers and autonomous agents
Add the AI browser, browser agent and general agent categories. Remove the apps from devices with MDM at the same time.
- Day 22 onward
Block the full list, allow by exception
Move to the complete list. Put approved tools on an allow rule above the block, owned by a named person and reviewed each quarter.
What to block, what to monitor, what to allow
A blocklist works well with a short written decision for each group of tools. This is the split most teams settle on.
Block
- AI browsers and browser agents
- Agent extensions that read every page
- Hosted MCP servers and MCP directories
- General autonomous agents with web access
- Unapproved coding agents on servers
Monitor
- Deep-research agents used by analysts
- Meeting and voice agents, until a policy exists
- Agent-building platforms used by IT itself
- Moderate and Low grade entries in the first weeks
Allow
- The chat assistant your company licensed
- The coding agent your engineering team approved
- MCP servers you host and review yourself
- Each exception with an owner and an end date
Firewall plus endpoint: the full picture
The firewall stops the conversation between the agent and its service. Device policies stop the agent from being installed. Use both and each one covers the other's blind spots.
At the firewall
- Covers every device on the network, managed or not
- Stops agents that run in the cloud for a user
- Logs who tried to reach which agent
- Updates in minutes when the list changes
On the device
- Removes AI browsers and blocks agent extensions
- Works off the office network
- Controls local MCP servers and coding agents
- See blocking AI browser extensions and blocking AI browsers
Give your firewall the list it is missing
One subscription covers the firewall, the DNS filter and the device policies. The lists update every day, and we set up your download access within one business day of your order.
- Plain domain list and firewall external list format
- CSV and JSON with category and risk grade
- DNS filter import files, hosts file, PAC file
- Extension and app identifiers for MDM and browser policy
- Updated every day
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
Firewall and AI: questions admins ask
Can a firewall block AI traffic?
Yes, for anything that crosses the network. A firewall can block the domains that AI browsers, agents, agent extensions, remote MCP servers and coding agents call.
It cannot see what runs only inside a device, such as a local MCP server, so pair the firewall list with browser and device policies.
Do I need TLS inspection to block AI on the firewall?
No. Blocking by domain works from DNS queries and the server name a browser sends when it opens a connection. TLS inspection is only needed to block a path inside a site you otherwise allow, or to see full URLs in logs.
Will blocking AI agent domains break ChatGPT, Copilot or Gemini?
Only if you block those domains. Every entry has a category and a risk grade, so you can block AI browsers and autonomous agents while you keep approved chat assistants on an allow rule above the block rule.
How do I block AI on a Palo Alto or Sophos firewall?
On a Palo Alto Networks firewall, add our list as an external dynamic list and use it in a URL Filtering profile with the action block. See the Palo Alto guide.
On Sophos Firewall, put the domains in a URL group and block that group in a web policy rule. See the Sophos, WatchGuard and UniFi guide.
How often does the firewall list change?
The list is updated every day. Firewalls that read a list from a URL pick up the change on their next refresh. For firewalls that need a manual import, most teams replace the list once a week.
Can I use the list next to my firewall's own web categories?
Yes. Our list is added as a custom or external list. It sits next to the categories you already use and does not replace or change them.
Can a firewall block AI browser extensions?
A firewall cannot see which extension is installed, but it can block the services an agent extension calls, which stops the extension from working. To remove the extension itself, use the extension block list in Group Policy, Intune or the Google Admin console.
Perplexity and Comet are trademarks of Perplexity AI, Inc. ChatGPT, ChatGPT Atlas and Codex are trademarks of OpenAI. Claude and Claude Code are trademarks of Anthropic, PBC. Cursor is a trademark of Anysphere, Inc.
Dia is a trademark of The Browser Company of New York. Opera Neon is a trademark of Opera Norway AS. Manus, Genspark, Pi-hole and AdGuard are trademarks of their respective owners. Gartner is a trademark of Gartner, Inc.
Gemini, Chrome and Google Admin console are trademarks of Google LLC. Microsoft, Copilot, Edge and Intune are trademarks of Microsoft Corporation. Fortinet and FortiGate are trademarks of Fortinet, Inc. Palo Alto Networks is a trademark of Palo Alto Networks, Inc.
Zscaler is a trademark of Zscaler, Inc. Cisco and Umbrella are trademarks of Cisco Systems, Inc. Sophos is a trademark of Sophos Ltd. WatchGuard is a trademark of WatchGuard Technologies, Inc. UniFi is a trademark of Ubiquiti Inc.
Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.