ChatGPT agent mode

ChatGPT agent mode at work: risks and how to block it

Agent mode itself was retired in 2026. Its abilities did not go away: they moved into ChatGPT Work, a cloud browser, and a desktop app that can use the browser and the screen. If staff use personal accounts, none of your workspace controls apply.

Jan 2025Operator, a browser-using agent, appears as a separate product
17 Jul 2025ChatGPT agent combines Operator and deep research inside ChatGPT
9 Jul 2026ChatGPT Work launches with a redesigned desktop app
Aug 2026Agent mode no longer offered; the standalone Atlas browser stops working
NowWork Cloud, cloud browser, desktop Computer Use and the Chrome extension

Next step

AI Agent Blocklist: block ChatGPT's agent surfaces where they are not approved

Extension IDs, desktop app identifiers and domains for ChatGPT and every other agent, ready for browser policy, MDM, DNS and firewalls.

From $99 a month for one organisation

ChatGPT agent removed? What actually changed

Many admins first notice the change through help desk tickets: the agent option vanished, or scheduled agent tasks went quiet. Here is what changed and what it means for your controls.

Removed

The agent mode option is gone

OpenAI's help center now says ChatGPT agent is no longer available. Users reported the option disappearing in early August 2026.

Stopped

Old agent tasks stopped running

Users reported that recurring tasks built on agent mode stopped producing results. Rebuild needed tasks in an approved tool, under review.

Disabled

Off for one role, on for another

In managed workspaces, role permissions can switch agent features off. Custom roles combine, so one role set to On can outweigh another set to Off.

The point for security teams: the risk did not leave with the name. The successor features can browse, use apps and operate the computer, so treat them as agents in your policy and inventory.

Where ChatGPT's agent abilities live now

Four surfaces carry the agent work today. Each one reaches different data and is controlled in a different place. Availability depends on plan, region and rollout.

Work Cloud and the cloud browser

  • Runs on OpenAI-managed infrastructure
  • The cloud browser cannot see local tabs, extensions, history or saved passwords
  • Website sign-in through it is not available in Enterprise or Edu workspaces
  • By default it asks before visiting a new site

Computer Use in the desktop app

  • Sees and operates native apps on Mac and Windows
  • macOS needs Screen Recording and Accessibility permission
  • On Windows it works on the active, visible desktop
  • Cannot approve OS security prompts or sign in as an administrator

The desktop app's built-in browser

  • Uses a separate browser profile
  • Managed Browser Use policies restrict sites and capabilities
  • Cannot automate file uploads

The ChatGPT Chrome extension

  • Works with existing tabs and signed-in accounts
  • Managed Browser Use policies apply for workspace users
  • Installs per browser profile, so personal profiles bypass workspace rules

Cloud and local work are separate. A cloud environment does not inherit a computer's files, apps, browser sessions or network access by default. Local work uses approved resources on a member's computer, and that computer must be online.

Data residency: OpenAI documents that if a workspace enforces data residency in any cloud policy, Allow local computer access is disabled for Work. If you rely on residency, confirm the setting before a pilot.

One gap to know about: OpenAI's documentation says that when Computer Use operates a browser directly as a native app, the Browser Use site and capability policies do not apply. Native-app policies and OS permissions do. Set both.

Is ChatGPT agent mode safe for company data?

The question now covers all the agent features above. The answer depends less on the product than on which account a person is signed in with and what that session can reach.

Signed-in tabs

Through the extension the agent works in tabs where staff are already logged in to mail, CRM and finance tools.

Everything on screen

Computer Use sees what the user sees. Confidential windows left open become part of the task.

Instructions in content

Pages, mail and documents the agent reads can carry hidden instructions. See prompt injection in AI browsers.

Personal accounts

A personal plan on a work laptop is outside every workspace control: no admin console, no compliance records.

Data retention

OpenAI states that local computer access with Work Cloud does not provide strict zero data retention. Check this against your requirements.

Connected apps

Plugins and apps reach source systems with the connected account's rights. Enterprise and Edu start with most new ones off; Business starts with them on.

Signs that ChatGPT's agent features are in use

You can usually tell within a day whether staff use these features, and whether they use your workspace or their own accounts.

SignWhere you see itWhat it tells you
ChatGPT extension in managed browsersGoogle Admin apps and extensions report, Defender browser extension inventoryAn agent can act in signed-in tabs on that profile
ChatGPT desktop app on laptopsIntune discovered apps, Mac MDM inventoryComputer Use and the built-in browser may be available
Screen Recording or Accessibility granted to the appmacOS privacy settings, MDM privacy reportsComputer Use has been switched on by the user
ChatGPT traffic from users outside your workspaceWorkspace member list compared with DNS or proxy logsPersonal accounts in use on company devices
Leftover Atlas installsMac inventory by bundle IDAn unsupported browser that should be removed

Phone

Started from a phone

Tasks can be scheduled from the mobile app and followed in the browser or desktop app. Mobile use reaches the same connected apps.

PC

Running on a laptop

Computer Use works on the device it is installed on. Approve it only on machines where the screen holds nothing it should not see.

Cloud

Running elsewhere

Work Cloud keeps going while the laptop is closed. Its reach comes from connected apps and any local access you allowed.

Before you approve ChatGPT's agent features for a team

Some teams will benefit. Settle these questions first, write the answers down, and review them when OpenAI changes the product, which in this area has been often.

Scope

  • Which roles get Work, cloud browser, Browser Use and Computer Use?
  • Are personal accounts blocked on the same devices?
  • Do new plugins stay off until reviewed?

Data

  • Which connected apps hold regulated data, and are they read-only?
  • Does your retention policy accept the documented limits?
  • Who pulls compliance records, and how often?

Actions

  • Is Always allow for websites removed?
  • Which actions must a person confirm?
  • Which native apps may Computer Use operate, if any?

Upkeep

  • Who owns the workspace settings?
  • Who tests managed configuration after each app update?
  • When is the approval reviewed again?

Schools and regulated teams: where staff handle pupil, patient or client records, keep Computer Use and the extension off unless a review has approved them, and block personal accounts first. Our page on students and AI agents covers the school side.

ChatGPT agent mode in Enterprise, Edu and Business: the controls

If ChatGPT is approved, govern its agent features inside your workspace. These are the controls OpenAI documents for workspace owners; options vary by plan and rollout, so check the current admin console.

ControlWhereWhat it does
Agent SecurityAdmin ConsoleThe area for agent policies. OpenAI describes it as replacing the older Policies and Config area.
Allow local computer accessAdmin Console > Permissions, under Work CloudLets Work Cloud use approved resources on a member's computer. Turning it off interrupts running turns.
Cloud computer capabilitiesAdmin Console > Permissions & roles > Workspace capabilitiesControls cloud browser use and cloud network access.
Custom rolesPermissions & rolesEach permission is Default, On or Off per role. Roles add up, so check every role a user holds.
Website approvalCloud browser settingsUsers choose Always ask, Auto approve or Always allow. Admins can remove Always allow for the workspace.
Plugin controlsPlugins and connectionsDecide which apps are available, to which roles, and where supported, read-only or limited actions.
Computer HistoryWorkspace Settings > Permissions & rolesOff by default for Business and Enterprise; granted per role before members opt in.
Compliance recordsCompliance APIRecords cloud orchestration events when local computer access with Work Cloud is used.

Managed configuration on devices

For workspace users, OpenAI documents a managed requirements file that users cannot override. It can be delivered as workspace-managed configuration, a system-level file, or through macOS MDM, and MDM requirements rank above Agent Security.

# requirements.toml: enforced settings for managed ChatGPT desktop clients
features.computer_use = false
features.browser_use = false
features.browser_use_external = false
features.in_app_browser = false
allow_appshots = false

Changes delivered by macOS MDM apply at the next app launch. Restart the app on a test device and confirm the features are gone before rolling out.

How we solve it: block ChatGPT's agents where they are not approved

Workspace controls only reach people signed in to your workspace. For personal accounts, unmanaged profiles and organisations that do not allow ChatGPT at all, block on the browser, the device and the network.

ChatGPT approved

  • Keep the workspace, govern features in Agent Security
  • Push managed configuration to devices
  • Block other agents with our list

Personal use only

  • Block the ChatGPT extensions by ID
  • Block the desktop app by package and bundle ID
  • Remove leftover Atlas installs

Not allowed at all

  • Everything in the middle column
  • Block chatgpt.com and openai.com at DNS and the firewall
  • Expect wide impact on other OpenAI services

Identifiers from our list for ChatGPT's agent surfaces

SurfaceIdentifierTypeWhere it goes
ChatGPT extension for ChromehehggadaopoacecdllhhajmbjkdcmajgChrome extension IDExtensionInstallBlocklist in Chrome policy or Google Admin
ChatGPT extension for EdgeodlomjlbamekndcpllcnffbgeohgkmjhEdge extension IDExtensionInstallBlocklist in Edge policy
ChatGPT desktop app, WindowsOpenAI.ChatGPT, 9PLM9XGG6VKSwinget ID, Microsoft Store IDIntune app rules, Store restrictions
ChatGPT desktop app, macOSchatgpt (Homebrew cask), process ChatGPTPackage and process nameJamf, Kandji or Mosyle rules
Leftover Atlas installscom.openai.atlasmacOS bundle IDMac MDM inventory and removal

Steps for a Windows fleet with Edge and Chrome

  1. Edge: in Group Policy open Administrative Templates > Microsoft Edge > Extensions and set Control which extensions cannot be installed with the Edge ID. The registry key is SOFTWARE\Policies\Microsoft\Edge\ExtensionInstallBlocklist. Blocked extensions already installed are disabled. Details on our Edge extension blocklist page.
  2. Chrome: set the Chrome extension blocklist policy with the Chrome ID, through Group Policy, Intune or the Google Admin console. See blocking Chrome extensions with Group Policy.
  3. Desktop app: block or uninstall the app by package ID with Intune, and add the process name to your endpoint rules.
  4. Network: if ChatGPT is not approved at all, add the domains to your DNS filter. Our DNS blocking guide covers resolvers and filters.
  5. Everything else: load the full agent list so other agent extensions and AI browsers do not fill the gap. See how to block AI agents.

Two related pages: the extension on its own is covered in ChatGPT Chrome extension security, and removing the retired browser in block ChatGPT Atlas. For OpenAI's coding agent, see block Codex CLI.

Check that the block holds

Test on one managed device per platform before you roll out. These checks take about fifteen minutes.

Browser

Extensions

The ChatGPT extension cannot be installed from the store, and an existing copy shows as disabled by your organisation.

Device

Desktop app

The app is removed or cannot start, and the managed configuration shows Computer Use and Browser Use as off for approved users.

Network

Domains

Where ChatGPT is not approved, the domains return your block page or a refused lookup on office Wi-Fi and VPN.

Agents that operate the screen raise their own questions about permissions and dedicated machines. Read computer use agent security before approving Computer Use for any team.

Repeat the checks after each major ChatGPT app update and after any change to your workspace roles. Product names and settings in this area have changed several times in a year, so a test that passed in spring proves little in autumn.

One list for ChatGPT and every other agent

  • 2,600+ AI agent domains and 5,300+ device identifiers, each with a risk grade
  • Chrome and Edge extension block lists for Group Policy, Intune and Google Admin
  • macOS and Windows identifiers for Jamf and Intune, plus DNS and firewall formats
  • Updated every day; download access set up within one business day of your order

See everything in the download on the AI Agent Blocklist page.

From $99 a month for one organisation See plans Ask a question

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.

ChatGPT agent mode: questions admins ask

Was ChatGPT agent mode removed?

Yes. OpenAI's help center now says ChatGPT agent is no longer available, and users reported the removal in early August 2026. Multi-step work moved to ChatGPT Work, launched on 9 July 2026, and browser tasks to the cloud browser and the desktop app.

Why did ChatGPT agent mode stop or keep stopping?

If the option disappeared or scheduled agent tasks stopped producing results, the likely reason is the retirement of agent mode in 2026. In a managed workspace, a role setting can also switch agent features off. Check with your workspace admin.

Is ChatGPT agent mode safe for company use?

The agent features act with real access: in the cloud, on the screen and inside logged-in Chrome tabs. In a managed workspace admins can limit them. Personal accounts on company devices are outside those controls, so many organisations block them.

Can admins disable ChatGPT agent features in Enterprise?

Yes, for workspace users: Agent Security, the Work Cloud and cloud computer capability settings, role permissions, and managed configuration keys such as features.computer_use. Options depend on plan and rollout, so check the current admin console.

Does blocking chatgpt.com stop the agent?

It stops ChatGPT on networks you filter, including approved workspace use. It does not remove the desktop app or the extension and does not apply off your network. Combine it with extension and app blocks.

What replaced ChatGPT Atlas agent mode?

OpenAI announced in July 2026 that it would discontinue the standalone Atlas browser, with deprecation targeted for 9 August 2026. Its browsing features moved into the desktop app and a Chrome extension. Remove leftover installs by bundle ID.

Is there a ChatGPT agent CLI or API?

OpenAI's command-line agent is Codex CLI, a coding agent with its own package and process names, governed separately. Developers also build agents on OpenAI's APIs, which reach openai.com services rather than the ChatGPT app.

Can ChatGPT control my PC?

The desktop app has a Computer Use feature. macOS needs Screen Recording and Accessibility permission; on Windows it works on the active, visible desktop. It cannot approve OS security prompts or sign in as an administrator, and admins can turn it off.

OpenAI, ChatGPT, ChatGPT Atlas, Operator and Codex are trademarks of OpenAI. Microsoft, Windows, Edge and Intune are trademarks of Microsoft Corporation. Google and Chrome are trademarks of Google LLC. Apple and macOS are trademarks of Apple Inc.

Jamf, Kandji and Mosyle 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.