AI browser risks
AI Browser Security Risks for Companies
AI browsers read your company's pages, send what they read to a cloud service and act inside the accounts your staff are signed in to. Most were installed without IT knowing.
This page is the risk register: what can go wrong, who is exposed, how serious it is and what to do in the next 30 days.
Our rating, company with no controls
A judgement for a typical office, not a measurement. Your own ratings depend on your data and staff.
AI Agent Blocklist
Removes the biggest risk on this page, unmanaged AI browsers and agents, with daily lists of their domains, extension IDs and app identifiers for your firewall, DNS filter, MDM and browser policies.
From $99 a month for one organisation
Are agentic browsers safe?
The short answer depends on what the agent can reach
An AI browser is as risky as the accounts and data open inside it. The same tool that is harmless for reading public news can be a serious risk in a finance team's browser.
Lower risk
- Public research on news and product sites
- No signed-in work accounts in that profile
- Managed business edition with admin controls
- The agent reads but does not click or type
Needs controls
- Intranet, wiki and CRM pages open in the same browser
- Agent may fill forms or send messages
- Staff use personal accounts on work devices
- No log of what the agent did
Block
- Mail, banking, payroll and HR systems in reach
- Cloud admin consoles and identity portals open
- Regulated data: health, card, student records
- Consumer AI browsers that IT never approved
The register
AI browser risk register
Use these rows as a starting point for your own register. Ratings assume an office with AI browsers installed by staff and no specific controls in place.
| Risk | How it happens | Who is exposed | Likelihood | Impact | Main control |
|---|---|---|---|---|---|
| Agent steered by page content | Instructions hidden in pages, images, links or mail are read as commands. | Anyone who asks the agent to read outside content | High | High | Block unapproved agents; approved agent only on allowed sites |
| Data sent to cloud back ends | Page text, open tabs, screenshots and form content go to the vendor's servers for processing. | Staff who browse internal apps with the assistant open | High | High | Block back-end domains; contract terms for the approved tool |
| Actions in signed-in accounts | The agent clicks and types where the user is logged in. Systems record the user, not the agent. | Users of mail, CRM, banking and admin portals | Medium | High | Separate agent profile; block agents on sensitive sites |
| Wrong actions and purchases | The agent misreads a task and orders, books, replies or deletes the wrong thing. | Procurement, assistants, support teams | Medium | Medium | Confirmations on; no payment details in agent profile |
| Memories and history kept | Summaries of visited pages and chats are stored by the vendor and reused later. | Legal, HR, anyone handling personal data | Medium | Medium | Memory off in approved tool; retention terms checked |
| Installs outside IT | Browsers installed in a user's own folders may need no admin rights and can be missed by inventory. | Every managed laptop | High | Medium | Block and remove by bundle ID, package and process name |
| Agent extensions | An extension adds agent features to Chrome or Edge and reads every page. | Users of managed and unmanaged browsers | High | Medium | Extension blocklist or allowlist policy |
| Password and autofill exposure | The agent works in a browser where passwords and cards autofill. | Finance, admins, executives | Low | High | No password manager or saved cards in agent profile |
| Audit and compliance gaps | No record shows which actions a person took and which the agent took. | Regulated teams, auditors, data protection | Medium | Medium | Approved tool with audit logs; policy on agent use |
| Training done by agents | Staff let an agent complete mandatory courses or quizzes. | Compliance and HR owners | Medium | Low | Block agents on the learning platform; policy wording |
Likelihood and impact are our qualitative view for planning, not measured rates.
Heat map
Where each risk sits
Start in the top right corner. Those risks are both likely and serious, and all of them shrink sharply when unapproved AI browsers cannot run.
The problem in detail
The risks behind the register
Each card explains how the risk shows up in a normal working day and one sign that it is already happening.
Prompt injection
Text planted on a page can redirect the agent. It is the most discussed risk and has its own page: how web pages steer agentic browsers.
Content leaves the company
To answer "summarise this", the browser sends the page to its AI service. If that page is a contract, a customer record or a payroll screen, the content now sits with a third party.
Sign: DNS lookups for AI browser back ends from finance or HR devices.
The agent is the user
Websites cannot tell the agent from the person. A message sent, a file shared or a setting changed by the agent appears in logs under the employee's name.
Sign: staff dispute actions that the logs say they took.
Wrong but confident
Agents misread dates, quantities and names. Gartner's advice lists erroneous actions as a reason to wait. A wrong order or a reply to the wrong customer is the everyday version.
Sign: orders or bookings nobody remembers approving.
Memories you did not plan for
Some AI browsers keep memories of what the user browsed to personalise later answers. That creates a store of business context outside your retention rules.
Sign: the assistant recalls internal projects in later chats.
Installs IT never saw
Like other browsers, AI browsers can often be installed by the user, then offer to become the default browser and to import bookmarks, history and passwords from the browser already in use.
Sign: a new default browser on managed laptops.
Agents as extensions
Agent extensions bring the same powers into your managed Chrome or Edge. They need permission to read and change every site. See how to judge AI extension risk.
Sign: agent extensions in your browser management reports.
Compliance gaps
Data protection records, vendor contracts and sector rules assume you know where data goes. A browser that sends pages to an unlisted processor breaks that assumption.
Sign: an AI vendor that is missing from your processor list.
Attestations by agent
Gartner warned that staff could hand mandatory training to an agent. Completion records then prove nothing, which matters in audits and after incidents.
Sign: courses completed in a fraction of the usual time.
Gartner AI browser risks
What Gartner advised and what it means for you
In December 2025 Gartner published advice titled "Cybersecurity Must Block AI Browsers for Now". It told security leaders to block AI browsers for the time being, and said organisations with low risk tolerance may need to keep them blocked for longer.
The reasons named in the coverage match this register:
- Indirect prompt injection that leads to rogue agent actions
- Erroneous actions caused by flawed reasoning
- Sensitive data from the assistant sidebar reaching AI back ends
- Staff using agents to complete mandatory training
Our page on Gartner's advice to block AI browsers turns it into steps for IT teams.
Browser by browser
The questions people ask about specific AI browsers
The risk classes above apply to every AI browser. What changes between products is the edition, the admin controls and how the agent reaches your data.
Perplexity launched Comet Enterprise in March 2026 with admin controls such as domain blocks, approvals, task limits and audit logs. A free copy installed by an employee is not covered by those controls.
Our list includes Comet's identifiers, for example ai.perplexity.comet on macOS and Perplexity.Comet on Windows. Detail: Comet security risks.
OpenAI retired the standalone ChatGPT Atlas browser in August 2026 and moved its agent browsing into ChatGPT. Old installs can remain on Macs with their local profile data.
Look for com.openai.atlas in your inventory and remove it. The agent features now arrive through ChatGPT, so review those too. Detail: Atlas security concerns.
Brave's agentic mode is opt-in and runs in a separate profile, away from the user's everyday cookies and sign-ins. A second model checks each action. Brave says these steps lower but do not remove prompt injection risk.
Brave also has an AI assistant, Leo, which admins can manage by policy. See AI features in Brave and Firefox.
These newer agentic browsers follow the same pattern: page content goes to an AI service and the agent can act for the user. Treat each one as unapproved until you have reviewed it.
Our list covers them, for example BrowserOS.BrowserOS on Windows. Detail: are these AI browsers safe for work.
Who is exposed
Teams that carry most of the risk
Finance
Banking portals, supplier records and payment approvals sit one tab away from any page the agent reads.
Executives and assistants
Mail, calendars and travel bookings are exactly the tasks AI browsers advertise.
IT and cloud admins
An agent in an admin console holds the keys to every other system.
HR and recruiting
Candidate files, salaries and health notes are personal data with strict rules.
Sales and support
CRM records and customer tickets are read and answered in volume, often with agents.
Developers
Code hosting, cloud consoles and secrets in web dashboards are all browser based.
Legal and compliance
Privileged documents sent to an AI service can create disclosure questions.
Schools
Student records, grading systems and online tests are open to agents on staff and student devices.
Signs to look for
How to tell AI browsers are already in use
- DNS or proxy logs show AI browser back ends such as perplexity.ai, chatgpt.com, genspark.ai or browseros.com from devices that never used them before.
- Software inventory lists unknown browsers installed in user folders rather than in Program Files or the system Applications folder.
- The default browser on managed devices changed without a ticket.
- Browser management reports new extensions that request access to all sites.
- Sign-in alerts show the same account active from a new browser on the same device.
- Helpdesk tickets mention "the assistant did it" or ask how to connect mail to a browser.
How we solve it
A 30-day plan to bring AI browsers under control
Each phase uses the AI Agent Blocklist files in tools you already run. Check the current admin console for exact menu names, which change between releases.
Days 1 to 5
Find
- Match software inventory against our macOS and Windows identifiers
- Match extension reports against our extension IDs
- Run our domain list against a week of DNS logs
Days 6 to 10
Decide
- Block all, or block by risk grade
- Name any approved agent and its edition
- Write the rule into your AI agent use policy
Days 11 to 20
Block
- Import the domain list into firewall or DNS filter
- Push extension blocks through Intune, Group Policy or Google Admin
- Block or remove apps with our bundle IDs, process names and package IDs in Jamf or Intune
Days 21 to 25
Clean up
- Remove existing installs with your MDM
- Ask users to sign out of sessions used in AI browsers
- Follow our removal steps
Days 26 to 30
Review
- Check DNS logs for blocked lookups
- Handle exception requests with an owner and end date
- Let the daily updates cover new tools
For a structured review of a tool someone wants to approve, use our AI agent risk assessment checklist.
Your policy
What to block, what to monitor, what to allow
| Block | Monitor | Allow |
|---|---|---|
| Consumer AI browsers on managed devices | Lookups for agent domains from guest and BYOD networks | One business edition agent with admin controls on |
| Agent extensions outside your allowlist | Extension installs reported by browser management | Chat assistants that cannot act in the browser |
| Higher-risk grades in our list, if you block selectively | Lower-risk grades you chose to leave open | Read-only research on public sites |
| Agents on finance, HR, identity and admin sites | Audit logs from the approved agent | Named users who completed agent training |
| Leftover installs of retired AI browsers | New AI browsers added in our daily updates | Time-limited exceptions with an owner |
The product
Shrink the register with one daily list
The AI Agent Blocklist lists 2,600+ AI agent domains and 5,300+ agent identifiers on devices: AI browsers, browser agents, agent extensions, desktop agent apps and coding agents. Each entry has the tool name, category, identifier type and a risk grade.
Every entry is checked before it enters the list, and the lists update every day. After an order we set up your download access within one business 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
Questions
AI browser security: questions from IT teams
Are agentic browsers safe for work?
Not by default. An agentic browser reads company pages, sends content to its vendor's cloud and acts inside accounts where the user is signed in. Some companies allow one managed agent for low-risk tasks, but unmanaged AI browsers on work devices are a risk most security teams do not accept.
What are the main security risks of AI browsers?
Prompt injection that steers the agent, company data sent to cloud back ends, actions taken in the user's signed-in accounts, wrong purchases or messages, stored memories and history, installs outside IT control, agent extensions with broad access and gaps in audit trails.
What did Gartner say about AI browser risks?
In December 2025 Gartner published advice titled Cybersecurity Must Block AI Browsers for Now. It recommended blocking AI browsers for the time being because of risks such as rogue agent actions caused by indirect prompt injection, erroneous actions and sensitive data reaching AI back ends.
Is the Comet AI browser secure for company use?
It depends on the edition. Perplexity launched Comet Enterprise in March 2026 with admin controls such as domain blocks, approvals, task limits and audit logs. A consumer copy installed by an employee has none of them, so block it and decide on the managed edition separately.
Is OpenAI's AI browser still a security risk?
OpenAI retired the standalone ChatGPT Atlas browser in August 2026 and moved its agent features into ChatGPT. Old installs and their local data can still sit on devices, and the agent features now arrive through ChatGPT instead, so the risk moved rather than disappeared.
Is Brave's AI browser secure?
Brave's agentic mode is opt-in, runs in a separate profile with its own cookies and uses a second model to check actions. Brave says these safeguards do not remove prompt injection risk. Companies can manage Brave's AI features with its group policies.
Is the Genspark AI browser safe?
Genspark's browser is an agentic browser, so the same risk classes apply: page content goes to a cloud service and the agent can act for the user. Treat it like any other AI browser that IT has not approved and block it until you have reviewed it.
Perplexity and Comet are trademarks of Perplexity AI, Inc. OpenAI, ChatGPT and Atlas are trademarks of OpenAI. Brave and Leo are trademarks of their owner. Genspark, Fellou and BrowserOS are trademarks of their respective owners.
Google, Chrome and Google Admin are trademarks of Google LLC. Microsoft, Edge, Intune and Windows are trademarks of Microsoft Corporation. Jamf is a trademark of JAMF Software, LLC. Apple and macOS are trademarks of Apple Inc. Gartner is a trademark of Gartner, Inc.
Names are used only to identify the products and publications discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.