Computer use agent security: AI that controls the screen
A computer use agent looks at the screen, decides, and clicks and types like a person. It needs no API and no integration, so it can work in any app your staff can open, with their logins and their data in view.
That flexibility is the risk. Here is what to watch, which permissions to lock and how to keep these agents to the machines you choose.
AI Agent Blocklist: keep computer use agents off machines that were not built for them
Desktop agents, browser agents, agent extensions and the platforms that run them, as device identifiers and domains for MDM, browser policy and firewall.
From $99 a month for one organisation
What computer use agents are and where they come from
The big AI labs and Microsoft all ship computer use now, as developer tools, inside assistants and inside agent builders. Open-source projects and agent platforms add more every month. These are the ones admins meet most.
A developer tool: the model returns screenshot, click and type actions, and the developer's own environment runs them. Public beta from October 2024.
The model behind Operator, OpenAI's first browser agent. Its capabilities later moved into ChatGPT's agent features.
Released in public preview on 7 October 2025 through the Gemini API and Vertex AI. Built mainly for browsers, with a safety check on each step.
Lets agents operate websites and Windows desktop apps on a machine you set up. By default it may act on any site or app there.
Operates native apps on the user's own Mac or Windows PC for ChatGPT Work tasks, inside the limits an admin sets.
An experimental agent for Windows Insiders that works through apps in a separate agent workspace under its own account.
Behind these sit agent platforms and open-source projects that let anyone build a screen or browser agent. A few example domains from our list:
Why screen-driving agents break your usual controls
Most security tools draw lines around apps, APIs or browser tabs. A computer use agent works at the level of pixels and keystrokes, the same layer as the person at the keyboard.
| Question | API agent | Browser agent | Computer use agent |
|---|---|---|---|
| What it sees | The data each API returns | Pages in its tab or browser profile | Everything visible on screen, from any app |
| What it can touch | Only the APIs it was granted | Websites, with the profile's logins | Any app, file dialog or setting the user can reach |
| Identity it acts with | A token you can scope and revoke | The browser session cookies | The signed-in user, in full |
| Controls that bind it | Token scopes, API gateways | Extension and site policies | OS permissions, device rules, dedicated machines |
| Where injected text hides | API responses | Web pages and emails | Anything rendered: pages, PDFs, chat windows, pop-ups |
A documented example: OpenAI states that its browser site and capability policies do not apply when its Computer Use feature operates a browser as a native app. Native-app policies and OS permissions apply instead. Screen-level agents need screen-level controls.
The six risks admins should plan for
Each card shows a rough severity for a typical office laptop with no special setup. A contained, dedicated machine lowers every one of them.
It sees everything
Screenshots capture open mail, chats, customer records and password manager pop-ups. With cloud models, that image usually goes to the provider with each step.
It acts as the user
Clicks and keystrokes carry the user's full rights. Sending, approving, deleting and paying look exactly like the employee did it.
Instructions on screen
Text in a page, document or image can tell the agent to do something else. Anthropic, Google and Microsoft all name this prompt injection risk in their own documentation.
Controls bypassed by design
No API means no token scope, no API gateway and no integration review. The agent uses the same front door as the person.
Credentials in reach
Agents that sign in to sites need passwords somewhere: typed by the user, stored in the tool, or visible in a vault window on screen.
Unattended runs
Background and scheduled runs mean nobody watches each step. Some agents can even work on a locked Mac for an active task.
Prompt injection is the risk that ties these together. Our page on prompt injection in AI browsers shows how hidden instructions reach an agent.
What vendors build in, and what stays your job
Safeguards vendors document
- Anthropic runs classifiers over screenshots that look for injected instructions and steer the model to check with the user.
- Google describes a per-step safety service that reviews each proposed action before it runs.
- OpenAI's Computer Use cannot approve OS security prompts, sign in as an administrator or drive terminal apps.
- Copilot Studio offers an allow list, HTTPS enforcement and email alerts to a named reviewer.
- Windows runs Copilot Actions in a separate agent workspace under its own account.
What no vendor setting does for you
- Decide which agents may run on which machines at all
- Stop staff installing an agent your contracts do not cover
- Keep agents off laptops with admin rights or production access
- Remove an agent's OS permissions on devices you manage
- Block the many smaller agent tools that have no admin console
Anthropic's own advice for its tool is a good baseline for any vendor: a dedicated virtual machine or container with minimal privileges, no access to account logins, internet limited to an allow list of domains, and a person confirming steps with real-world consequences.
Signs computer use agents are already on your devices
Look for the permissions these agents need and the helpers they install. On a Mac the two privacy permissions are the clearest tell.
How we solve it: a four-layer setup for computer use agents
Decide which machines may run screen-driving agents, then make every other machine refuse them. The AI Agent Blocklist supplies the identifiers for the refusing part; your MDM, browser policy and firewall enforce it.
- Deploy a Privacy Preferences Policy Control payload (com.apple.TCC.configuration-profile-policy) from your MDM to supervised Macs.
- Under Screen recording, deny the agent apps from our macOS bundle ID list. Apple documents this setting as a way to deny screen capture to named apps.
- Under Accessibility, deny the same apps by bundle ID, with the code requirement you read using codesign -dr - on a test Mac.
When several profiles conflict, macOS applies the more restrictive setting, so a deny profile holds even if another profile is looser. For the browser side on Macs, see blocking Chrome extensions with Jamf.
- Keep Experimental agentic features off and staff out of the local administrators group, since only an admin can turn it on.
- Use App Control for Business rules to stop desktop agent apps from running, using our Windows package IDs and process names to locate them.
- On machines approved for agents, allow only the apps the workflow needs. Microsoft gives the same advice for Copilot Studio machines.
Copilot's own agent switches are covered in disable Copilot actions, and Intune rollout steps in block AI with Intune.
- Add our agent extension IDs to the Chrome and Edge extension block lists, so browser agents cannot be added to work profiles.
- On agent machines, set a URL allow list so the agent's browser reaches only the sites in its job.
- Keep agent browsers in their own profile with no saved passwords for staff accounts.
Extension rollout in detail: block AI browser extensions. Vendor-specific browser agents: Claude in Chrome and ChatGPT agent features.
- Load our domain file into your firewall or DNS filter and block agent platforms and model APIs you have not approved.
- Allow the approved agent's domains only from the subnet of the agent machines.
- Alert on new agent domains in our daily updates hitting your network.
Formats for each firewall are on blocking AI agents at the firewall.
Vendor settings still matter: where your company licenses an agent, use its admin controls too.
For example, ChatGPT's managed configuration documents features.computer_use = false to turn native Computer Use off and computer_use.allow_locked_computer_use = false to stop users enabling Locked Use on managed Macs.
Block
- Desktop and browser agents on staff laptops
- Screen Recording and Accessibility for agent apps
- Agent platforms with no contract
Monitor
- Privacy permission changes on Macs
- Agent connectors on Windows pilots
- New computer use tools in Copilot Studio
Allow
- Approved agents on dedicated machines
- Least-privilege service accounts
- Site and app allow lists per workflow
Build sheet for a machine that may run a computer use agent
If a team has a real case for a screen-driving agent, give it a machine built for that job and nothing else. Use this sheet when you set one up, and attach it to the approval.
| Setting | What to choose | Why it matters |
|---|---|---|
| Machine | A dedicated virtual machine or physical device, used only by the agent | Nothing else on screen means nothing else to capture or click |
| Account | A service account with the least access the workflow needs | Mistakes and injected steps stay within a small blast radius |
| Admin rights | None for the agent account | The agent cannot install software or change system settings |
| Installed apps | Only the apps in the workflow, enforced by application control | The agent cannot wander into a terminal or an admin tool |
| Browser | One profile, a URL allow list, no saved staff passwords | Injected links lead nowhere useful |
| Credentials | Stored in a vault scoped to one login domain each | Secrets never sit in prompts, notes or screenshots |
| Network | Outbound only to the approved agent service and the workflow's sites | A misdirected agent cannot send data to a new destination |
| Approvals | A named person confirms payments, submissions and messages | Consequential steps get a human check |
| Records | Keep the agent's step log and screenshots for an agreed period | You can show what the agent did and why |
| Review date | A fixed date to renew or retire the approval | Pilots do not quietly become permanent |
Agents also need identities you can scope and revoke. Our page on agent identity and permissions covers service accounts and token scopes in more depth.
Add computer use agents to your computer use policy
Most companies already have an acceptable use policy for computers. It was written for people. These clauses extend it to software that uses the computer on a person's behalf.
Full policy wording for all agent types is in our AI agent acceptable use policy, and a scoring approach in the AI agent risk assessment.
Identifiers for the agents you do not want on the screen
Computer use is spreading from a few labs to many small tools. Our files track desktop agents, browser agents, agent extensions and agent platforms, with a risk grade on each entry so you can block all of them or the riskier ones first.
What each file contains is on the AI Agent Blocklist page.
See plans and prices- macOS bundle IDs for privacy and app blocking profiles
- Windows package IDs and process names for App Control and Intune
- Chrome, Edge and Firefox extension IDs of agent extensions
- 2,600+ AI agent domains for firewalls and DNS filters
- CSV and JSON with tool name, category, identifier type and risk grade
- Updated every day; access set up within one business day of ordering
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
Computer use agent questions
What is a computer use agent?
A computer use agent, often shortened to CUA, is an AI agent that operates a computer the way a person does. It takes screenshots, decides what to do, then moves the mouse, clicks and types. It works with any website or desktop app, even ones that have no API.
Are computer use agents safe to use at work?
They can be used safely only in a contained setup: a dedicated machine or virtual machine, a least-privilege account, an allow list of sites and apps, no stored personal logins and a person approving consequential steps.
On an ordinary staff laptop they see and can act on everything the user can.
Which permissions does a computer use agent need on a Mac?
To see the screen it needs Screen Recording, and to click and type it needs Accessibility. On managed Macs a Privacy Preferences Policy Control profile can deny screen capture to named apps and allow or deny Accessibility by bundle ID or path. The payload needs supervision when it comes from an MDM.
What is a computer use policy?
Traditionally it is the acceptable use policy for company computers. It now needs clauses for AI agents that operate those computers: which agents are approved, which accounts and data they may touch, when a person must approve an action, and who is accountable for what the agent does.
Can a computer use agent get around browser policies?
It can sidestep controls that only apply inside one channel. For example, OpenAI documents that its browser site policies do not apply when its Computer Use feature operates a browser as a native app. Screen-level agents need screen-level and device-level controls.
Do computer use agents work when the computer is locked?
Some can. OpenAI documents a macOS option called Locked Use that can temporarily unlock a Mac for an active task and relocks on local input, and a managed setting that stops users enabling it. Check each agent's documentation and set the managed option before rollout.
How do developers use a computer use agent API or SDK safely?
Run the agent in a dedicated container or virtual machine with no personal data, and limit its internet access to an allow list.
Keep secrets out of the environment and require human confirmation for purchases, sign-ups and terms. Anthropic's documentation for its computer use tool recommends these precautions.
Does blocking a domain stop a computer use agent?
Blocking the agent's own service and model API domains stops cloud-driven agents from working on your network. A locally installed agent also needs a device block on its app and process. Our files contain both the domains and the device identifiers.
Anthropic and Claude are trademarks of Anthropic, PBC. OpenAI, ChatGPT and Operator are trademarks of OpenAI. Google, Gemini, Vertex AI and Chrome are trademarks of Google LLC. Microsoft, Windows, Copilot, Copilot Studio, Edge and Intune are trademarks of the Microsoft group of companies.
Apple, Mac and macOS are trademarks of Apple Inc. Jamf is a trademark of Jamf Software, LLC. Other product names and domains shown belong to their owners. Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.