Gartner advisory, December 2025

Gartner Says Block AI Browsers: What IT Teams Should Do Now

Gartner told security leaders to block AI browsers for now. Below: what the advisory says, what changed since it came out, and the steps that turn it into rules on your network and devices this month.

Next step

AI Agent Blocklist

Turns the advice into rules: AI browser domains for your firewall and DNS filter, plus app and extension IDs for Intune, Jamf and Chrome policies, updated every day.

From $99 a month for one organisation

What Gartner's AI browser advisory says

The full text is for Gartner clients. The points below follow public reporting on it, in our own words. They describe a category of software, not one product.

Feature 1: the AI sidebar

A panel that summarises, searches, translates and talks about the page in front of the user. To do that, it may send page content, open tabs and browsing history to the AI service behind it.

Feature 2: agentic actions

An agent that completes tasks on websites by itself, often inside sessions where the user is already signed in. It clicks, types, fills forms and submits them.

The risks it names

1

Data leaves through the sidebar

Sensitive content on screen can reach the AI back end unless privacy and security settings are hardened and centrally managed.

2

Defaults favour convenience

Out of the box, settings lean towards a smooth user experience rather than security. Most users never change them.

3

Rogue actions from prompt injection

Instructions hidden in web content can steer the agent into actions the user never asked for.

4

Wrong actions from flawed reasoning

The agent can fill in a form wrongly, order the wrong office supplies or book the wrong flight.

5

Credentials on phishing sites

If the agent is led to a fake site, sign-in details can be lost or misused.

6

Training completed by agents

Staff could hand mandatory, repetitive tasks such as security awareness training to an agent.

What it recommends

  • Block users from downloading and installing AI browsers, using existing network and endpoint controls.
  • Assess the security of the AI service behind any AI browser before considering it.
  • If use is approved, teach users that anything on screen may be sent to the AI back end.
  • Ask users to close tabs with highly sensitive data while the sidebar is in use.
  • Limit what agents can do, for example keep them away from email.
  • Use settings that stop the AI browser from keeping data.
  • Run risk assessments before any deployment, and expect a long list of prohibited uses.
  • Monitor the AI browser fleet to enforce the policy, and accept that low risk tolerance may mean a longer block.

Source: public reporting on the Gartner advisory of early December 2025. For the full text, see Gartner directly.

Why agentic browsers are seen as too risky right now

The core issue is simple. An AI browser reads untrusted text from the web and acts with trusted access. Security tools built for humans clicking links do not see the difference.

Untrusted page

A web page, a comment, an email or a link contains text written for the AI, not for the reader.

Read as an order

The assistant cannot reliably tell the user's request apart from instructions inside the content it reads.

Acts as the user

The agent uses the user's signed-in sessions: mailbox, files, CRM, payment portals.

YearWhat researchers showed for AI browsers in generalWhat it means for IT
2025Hidden text in a public comment could steer an AI browser's assistant into reading a user's email address and a one-time code and posting them.Summarising a page was enough. No download, no macro, no warning.
2025A crafted link could carry instructions that gathered data from the user's connected email and calendar.Connectors to mail and calendars widen the blast radius of one click.
2025Text dressed up as a web address, pasted into an AI browser's address bar, could be handled as a trusted command.Even the address bar became an input for the agent.
2025A page could plant instructions in an AI assistant's memory that carried into later sessions.One bad page can affect work done days later.
2025Instructions hidden after the # in an ordinary web address could reach AI browser assistants, while the website never saw them.Trusted sites can carry hostile instructions without being hacked.

These are patterns for the whole category, not findings about one product. Makers describe prompt injection as an open problem for the industry. Our prompt injection explainer and AI browser security risks page go deeper.

What has changed since the advisory came out

Nearly a year on, the category has grown rather than shrunk. Agent features spread into mainstream browsers and desktop apps, and some vendors added admin controls.

March 2026

Comet Enterprise

Perplexity launched an enterprise edition of Comet with MDM deployment on macOS and Windows, over 500 Chromium policies, and admin choice over the Assistant and Agent.

March 2026

Comet on iPhone

Comet reached iOS on 18 March 2026, after Android in November 2025. Phones are now part of the AI browser question.

April 2026

Agents inside Chrome

Google opened Gemini auto browse in Chrome to eligible US Workspace users, governed by Chrome policies such as GeminiActOnWebSettings.

July 2026

Atlas retirement announced

On 9 July OpenAI announced the end of ChatGPT Atlas and launched a desktop app with a built-in browser and the ChatGPT Work agent.

August 2026

Atlas stops working

From 9 August Atlas stopped working, with no further security updates. Copies left on Macs should be removed.

October 2026

Dia on Windows

Reports in October said the Dia browser had reached Windows after months on macOS only, widening its reach in Windows fleets.

Does the advice still stand? For most organisations, yes. Enterprise editions now offer controls, which gives you options later. But the number of places an agent can live went up, so a default block plus named exceptions is still the safer start.

Details on two well-known AI browsers: Comet security risks and ChatGPT Atlas after retirement.

What the advisory means for your role

The same advice lands differently on each desk. Here is the short version for the people who usually own a piece of it.

CISO
  • Set the default: AI browsers blocked, exceptions by request.
  • Name an owner for each layer: network, endpoint, browser, identity.
  • Report coverage, not intentions: how many devices enforce the block.
IT and endpoint admin
  • Block install and launch by bundle ID, package ID and process name.
  • Find and remove existing copies, including retired Atlas.
  • Add agent extension IDs to ExtensionInstallBlocklist.
Network and firewall team
  • Load AI browser and agent domains as an external list.
  • Run it in log mode for a few days, then block.
  • Cover DNS for roaming laptops, not only the office firewall.
MSP
  • Offer the block as a standard control for every client.
  • Keep one list and push it to each client's firewall and MDM.
  • Record exceptions per client so audits are easy.
School and district IT
  • Agents that complete homework and quizzes are an academic integrity issue too.
  • Block AI browsers on student devices and agent extensions in Chrome.
  • Tell teachers which AI tools stay approved.
Compliance owner
  • Add AI browsers and agents to the acceptable use policy.
  • Ask for a risk assessment before any exception.
  • Check that agent actions can be traced to a person.

A 30/60/90 day plan to follow the advisory

Start with what stops the most risk for the least effort, then close the gaps. Every step uses controls you already own.

Days 1 to 30

Block and find

  1. Load the AI browser and agent domain list on the firewall and DNS filter.
  2. Search inventory for app IDs such as Perplexity.Comet, ai.perplexity.comet and com.openai.atlas.
  3. Add app rules in Intune App Control for Business, AppLocker or Jamf Restricted Software.
  4. Tell staff what changed and how to ask for an exception.
  5. Remove retired ChatGPT Atlas copies.
Days 31 to 60

Close the gaps

  1. Block agent extensions by ID in Chrome and Edge, for example fcoeoabgfenejglbffodgkkbkcdhcgfn and cecngibhkljoiafhjfmcgbmikfogdiko.
  2. Set Chrome and Edge AI policies such as GeminiSettings and EdgeCopilotEnabled per user group.
  3. Turn off agent and browser features in the AI services you approved.
  4. Extend DNS protection to laptops off the office network.
Days 61 to 90

Decide on exceptions

  1. Review who asked for an AI browser and why.
  2. Run a risk assessment of the AI back end for any candidate.
  3. Allow only enterprise editions with central policies, agent off by default.
  4. Write the result into your acceptable use policy.
  5. Set a review date to revisit the block.

Step-by-step controls for each layer are in our guide to blocking AI browsers. For the firewall part, see blocking AI agents at the firewall.

Block now or allow later: how to decide

"For now" is not "forever". The question is which conditions must be true before an AI browser earns a place on your devices.

Keep blocking if

  • You handle regulated, client or student data in the browser.
  • You cannot assess the AI service behind the browser.
  • The browser offers no central policies for its AI features.
  • Agents could reach email, payments or admin consoles.
  • You cannot see or log what the agent did.
  • Your risk tolerance is low, as the advisory warns.

Consider a managed exception when

  • An enterprise edition deploys through your MDM.
  • The agent is off by default and can be scoped to named sites.
  • Data retention can be turned off centrally.
  • Mail and file connectors can be disabled.
  • Sign-in uses company accounts you control.
  • A named team has a clear business need and an owner.
Where does data go?Which regions, how long it is kept, and whether it is used for training.
What can the agent touch?Tabs, cookies, saved passwords, local files, connected apps.
What can admins enforce?Policies for assistant, agent, extensions, sites and memory.
What is logged?Agent actions tied to a user, exportable to your tools.

A full question set is in our AI agent risk assessment guide, and policy wording is in the AI agent acceptable use policy.

From advice to rules: what each recommendation needs

Every recommendation needs a current list of what to block. That list is what the AI Agent Blocklist provides, in the format each tool reads.

RecommendationWhere you enforce itWhat our files give you
Block downloadsFirewall, DNS filter, proxyAI browser and agent domains, for example browseros.com and genspark.ai, as an external list, DNS import, hosts or PAC file
Block installsIntune, Group Policy, Jamf PromacOS bundle IDs, Windows package IDs and process names of AI browsers and agent apps
Limit agentsChrome and Edge policies, Google Admin consoleAgent extension IDs ready for block lists
Assess riskYour review processA risk grade on every entry, from Low to Severe, plus category
Monitor the fleetInventory and logsIDs and domains to search for, refreshed every day
Educate usersAcceptable use policyCategory names you can quote in the policy

Beyond browsers: AI agents, coding agents and failed projects

Searches for Gartner's view on AI agents often lead here. The browser advisory is one piece; the same pattern of an agent acting with a person's access shows up elsewhere too.

40%+of agentic AI projects will be canceled by the end of 2027, Gartner predicted in June 2025, citing escalating costs, unclear business value or inadequate risk controls.

AI agents

General agents connect to mail, files and business apps. The questions are the same: who approved it, what can it reach, who sees what it did. See AI agent security risks.

Coding agents

Coding agents run commands in a terminal with a developer's rights. We apply the same block, review, allow approach. See coding agent risks.

Governance

Weak risk controls are one reason projects get canceled. An inventory and clear approval rules come first. See AI agent governance.

Put the advisory into practice this week

The AI Agent Blocklist covers AI browsers, AI agents, agent extensions, coding agents and MCP servers. It holds 2,600+ AI agent domains and 5,300+ agent identifiers on devices, each with a category and a risk grade.

New AI browsers and agents are added as they appear, every day. After an order we set up your download access within one business day.

See plans

Read what is inside on the AI Agent Blocklist page, or browse the list of AI browsers.

  • Firewall external list format and plain domain list
  • DNS filter import files, hosts file and PAC file
  • Chrome and Edge extension block lists for Group Policy, Intune and the Google Admin console
  • macOS and Windows identifiers for Jamf and Intune
  • CSV and JSON with every column

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.

Gartner and AI browsers: common questions

What did Gartner say about AI browsers?

In early December 2025 Gartner published an advisory titled Cybersecurity Must Block AI Browsers for Now. It advises blocking AI browsers with existing network and endpoint controls, because the AI sidebar can send sensitive data to the AI back end and the agent can be steered into rogue or wrong actions.

Which AI browsers does the Gartner advisory cover?

It defines AI browsers by two features: an AI sidebar and an agent that acts on websites, often in signed-in sessions. It is about the category, not one product: any browser with both features fits the definition, whatever its maker.

Why does Gartner think agentic browsers are too risky?

Reported reasons: data sent from the sidebar to the AI back end, convenience-first defaults, prompt injection that triggers rogue actions, wrong actions from flawed reasoning, credentials lost on phishing sites, and agents completing mandatory training.

How long should we block AI browsers?

The advisory says for now. Reporting on it says the risks may not be fully eliminated, so low risk tolerance may mean a longer block. Set a review date, for example every six months, and check whether enterprise controls have matured.

Does the advice apply to Gemini in Chrome and Copilot in Edge?

The advisory is framed around browsers that combine an AI sidebar with agent actions. Mainstream browsers now add similar features, so many teams apply the same rules there with policies such as GeminiSettings, GeminiActOnWebSettings and EdgeCopilotEnabled.

Is ChatGPT Atlas still a concern now that it is retired?

Yes, in two ways. Atlas stopped working from 9 August 2026 and gets no security updates, so remove leftover copies. Its browsing and agent features moved into the ChatGPT desktop app and a Chrome extension, which need their own controls.

Do we need new security tools to follow the advisory?

Not necessarily. Reporting on the advisory says to block installs with existing network and endpoint controls. What most teams lack is a current list of what to block: AI browser domains, app identifiers and agent extension IDs that change every month.

What does Gartner say about agentic AI projects failing?

In June 2025 Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.

Gartner is a trademark of Gartner, Inc. Perplexity and Comet are trademarks of Perplexity AI, Inc. OpenAI, ChatGPT and ChatGPT Atlas are trademarks of OpenAI. Google, Chrome and Gemini are trademarks of Google LLC.

Microsoft, Windows, Edge, Copilot and Intune are trademarks of Microsoft Corporation. Apple, Mac, macOS and iPhone are trademarks of Apple Inc. Jamf and Jamf Pro are trademarks of Jamf. Dia, Opera Neon, Genspark, Fellou, BrowserOS and Claude 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 Gartner or by these companies.