Are AI Chrome extensions safe? How to judge the risk

Some are fine, some send more page content out than people expect, and some act on your staff's behalf in every signed-in app. You cannot tell which from the store rating, but you can tell from five or six plain questions.

Next step

AI Agent Blocklist: the IDs of AI and agent extensions, each with a risk grade, so you block the risky ones by policy

Chrome, Edge and Firefox extension IDs, tool names and categories, updated every day.

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

Can Chrome extensions be a security risk?

Yes. An extension runs inside the browser with whatever access it was granted at install, and most never misuse it. The problem is that you cannot see intent, and the access stays even when the code changes.

Reading pages

Access to all sites means access to webmail, CRM records, payroll and anything else open in a tab.

Taking sessions

Cookie and tab access can expose signed-in sessions, which attackers reuse without needing passwords.

Changing overnight

Updates install silently. A new owner, or someone who gained access to the developer's account, can ship new code to every user at once.

Copying trust

Lookalikes reuse the name and design of popular tools, sometimes with store badges, to collect data.

2024

Researchers reported that fake store emails gave outsiders access to several developers' store accounts, and altered updates reached dozens of extensions.

Millions

of people had those extensions installed when the altered versions went out, according to public reporting.

2025

Researchers reported lookalike AI sidebars that sent users' AI chat conversations and open tab addresses to an outside server.

Timed

uploads at a fixed interval were the pattern in that case, so steady, timed traffic from browsers to an unknown domain is worth checking.

Who is most exposed when an AI extension misbehaves

The same extension carries different risk depending on what the user can reach. Start your review with the people whose browsers hold the most sensitive sessions.

Finance

Banking portals, payment approvals and supplier records. An agent that can click can also approve.

HR and payroll

Personal data of every employee, salaries and health information sit one tab away.

Sales and support

CRM and ticket systems full of customer data, plus email that customers trust.

Developers and admins

Cloud consoles, code hosting and admin panels where one action changes production.

Executives

Board papers, deal rooms and inboxes that attackers target first.

Students

School accounts, learning platforms and minors' data, often on shared Chromebooks.

Contractors

Short-term staff on unmanaged browsers, signed in to your systems with fewer controls.

Remote staff

Mixed personal and work profiles on one machine, where personal extensions meet work sessions.

Permissions decoded: what each warning really allows

Chrome shows a short warning when an extension is installed. The wording is mild; the access is not. This table translates the most common warnings for AI extensions.

Warning or permissionWhat it allowsWhy AI tools ask for itExposure
Read and change all your data on all websitesRead and edit every page the user opensSummarise any page, fill forms, act on sitesHigh
Read and change your data on a number of websitesThe same, but only on listed sitesAdd features to one AI chat siteMedium
Read your browsing history, tabsSee addresses and titles of open and past pagesWork across tabs, pick contextHigh
Communicate with cooperating native applicationsTalk to a program installed on the computerConnect the browser to a desktop app or coding agentHigh
Manage your downloadsStart, read and open downloaded filesSave exports and generated filesMedium
Display notifications, storageShow alerts and keep local dataReminders and saved promptsLow
Admins can refuse permissions centrally. In the Google Admin console that lives under the additional app settings for Permissions and URLs; on Windows and Mac it is the blocked_permissions part of the ExtensionSettings policy. See blocking AI extensions in the Google Admin console and Group Policy and the registry.

How do I know if a Chrome extension is safe? Six questions

Use these questions for any extension a user requests. Each has a low risk answer and a high risk answer. Two or more high answers mean block, or allow only with an owner and a review date.

1

Who publishes it?

A company you can name, with a support address and a data policy, is easier to hold to account than an anonymous developer.

Known companyUnknown or new account
2

What does it ask to read?

Fewer sites and fewer permissions is better. An AI tool that needs every site for a narrow job deserves a second look.

One or two sitesAll sites, tabs, cookies
3

Does it only read, or also act?

Reading can expose data. Acting can move money, send emails and change records. Agent extensions do both.

Reads on requestClicks, types, submits
4

Where does the data go?

Content sent to an AI service is stored under that service's terms. Check retention, training use and the country of processing.

Your licensed serviceUnknown servers
5

How stable is it?

Frequent permission changes, a recent change of owner or a sudden rename are warning signs worth checking before you allow it.

Steady historyNew owner, new permissions
6

Is there a business need?

If the company already licenses a tool that does the job, a second extension adds risk without adding value.

Approved need, owner namedNice to have
For a fuller review template, including how to record decisions, see our AI agent risk assessment guide.

Why AI extensions carry more risk than ordinary ones

Two things change when an extension has AI inside: the content leaves the browser, and the extension can be talked into doing things. Agent extensions add a third: they act.

Ordinary extension
  • Code decides what happens
  • Data often stays local
  • Behaviour changes only with updates
  • Risk comes from the publisher
AI sidebar or helper
  • Sends page text and files to a model
  • Answers depend on page content
  • Hidden text can mislead it
  • Risk from publisher and AI service
AI agent extension
  • Clicks, types and submits
  • Uses the user's signed-in sessions
  • Any website can try to steer it
  • Risk from publisher, service and every page visited
The steering problem is called prompt injection. Our page on prompt injection in AI browsers shows how it works, and computer use agents covers agents that control the whole desktop.

Is the Claude Chrome extension safe?

Claude in Chrome is Anthropic's browser agent extension. It opens tabs, reads pages, clicks and types, and it shares the browser's login state, so it can reach any site the user is signed in to.

3

site categories blocked by default at launch: financial services, adult content and pirated content.

3

approval modes: Manually approve, Automatically approve and Skip all approvals. Your use policy should say which one staff may choose.

2

site lists for Team and Enterprise owners, an allowlist and a blocklist, under Organization settings > Claude in Chrome.

1

separate browser profile, without banking, health or government accounts, is how the maker advises using the extension.

What the maker built in

  • Site-level permissions that users can grant and revoke
  • Confirmation before high-risk actions such as publishing, purchasing or sharing personal data
  • Safety classifiers that look for suspicious instructions
  • Admin controls to turn the extension on or off and restrict websites with allowlists and blocklists

What stays your problem

  • It acts inside real, signed-in company sessions
  • A personal account sits outside your Team or Enterprise site lists
  • Users can widen its permissions for convenience
  • It can connect to a coding agent on the same machine
Our view for companies: treat any browser agent as you would an AI browser, which Gartner advised blocking in December 2025 in "Cybersecurity Must Block AI Browsers for Now". If you allow it, do it for named people, with site restrictions, after a pilot. Our page on blocking Claude in Chrome has the steps, and Gartner's advice on AI browsers explains the reasoning.

Store signals you cannot rely on

Users judge extensions by what the store shows. Each of these signals has failed in a real case.

It has a Featured badge.
Researchers reported in 2025 that a lookalike AI sidebar carried that badge while it collected chats.
Millions of users trust it.
Popular extensions are the most attractive targets for takeover, because one update reaches everyone.
The developer uses two-factor sign-in.
In the 2024 cases, access came through an approved app, which did not trigger a second factor.
The privacy policy says no personal data.
Policies describe intent, not behaviour. The 2025 lookalikes had policies that claimed exactly that.
It was fine when we checked last year.
Extensions update without asking. A review is a snapshot, not a guarantee.

Signs an extension has changed for the worse

You will rarely get a warning from the store. These signals, seen by users or in your logs, are reasons to block first and investigate second.

What users notice

  • Chrome asks to approve new permissions after an update
  • The extension's name, icon or publisher changes
  • New tabs, pop-ups or redirects appear on familiar sites
  • Search results start coming from a different engine
  • Forms are filled in or buttons clicked without the user acting

What IT sees

  • A spike in traffic from browsers to a domain nobody recognises
  • The same new extension version arriving on many devices at once
  • Requests at regular intervals, such as every half hour
  • Sign-ins to company apps from places the user has never been
  • Help desk tickets about an AI tool nobody approved
If you see these signs, block the extension ID first, then rotate sessions for affected users. Blocking by ID turns off installed copies on managed browsers within the next policy refresh.

Questions to ask before approving an AI extension

When a team asks for a specific AI extension, send the maker these questions. Clear, written answers make approval easy; vague ones make the decision for you.

  1. DataWhat page content, files and prompts leave the browser, and where are they processed and stored?
  2. TrainingIs customer content used to train models, and can a business account switch that off?
  3. ActionsCan the extension act on pages on its own, and can admins turn that off?
  4. Admin controlsCan we restrict sites, accounts and features centrally, and audit use?
  5. OwnershipWho owns the publisher account, and how are updates reviewed and signed off?
  6. IncidentsHow will you tell customers about a security issue, and how fast?

Ad blockers and other everyday extensions

Questions about ad blockers and small utility extensions come up in the same breath as AI. The same six questions apply, with one difference: their broad access is expected, so the publisher and history matter most.

Reasonable on company devices

  • One approved ad blocker from an established publisher
  • Password manager your company licenses
  • Tools force installed by IT with an owner

Worth blocking by default

  • Several overlapping blockers from unknown publishers
  • Free VPN and proxy extensions
  • Coupon, download and video helpers that read all sites
  • AI sidebars and agents outside your approved list

From judging to enforcing: turn the answer into policy

A risk review only helps if the browser follows it. These five steps take you from a decision to a policy that holds on every managed device.

  1. InventoryList installed extensions on enrolled browsers and match them by ID.
  2. ClassifyMark each as approved, limited or blocked, using the six questions.
  3. ImportLoad AI and agent extension IDs from our file into your block list.
  4. LimitBlock risky permissions and keep extensions off sensitive sites.
  5. RefreshApply the daily update so new agents and copies are covered.
fcoeoabgfenejglbffodgkkbkcdhcgfnClaude, Anthropic's browser agent extension
pninphncmkphpjgfffammoenglghcjiiPerplexity, the Perplexity extension
ofpnmcalabcbjgholdjcjblkibolbppbMonica, an AI assistant sidebar
camppjleccjaphfdbohjdohecfnoikecMerlin AI, an AI assistant sidebar

Four example rows, shown to illustrate the format. Listing a tool here is not a statement about its quality; it shows the kind of extension a company decides on centrally.

Building an approved list instead? Use the Chrome extension allowlist guide. Managing Macs with Jamf? See blocking extensions with Jamf.

Let the list keep up with new AI extensions

New AI sidebars, agents and copies appear every week. The AI Agent Blocklist tracks them and gives each a risk grade, so you can block all agent extensions or only the higher grades.

From $99 a month for one organisation. We set up download access within one business day of your order.

See plans Ask a question
  • Chrome, Edge and Firefox extension IDs of AI and agent extensions
  • Risk grade, category and tool name on every row
  • Policy files for Google Admin, Group Policy and Intune
  • AI agent domains for firewalls and DNS filters
  • CSV and JSON with all columns, 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

We use your details only to answer you. No newsletters.

Questions about extension safety

Can Chrome extensions be a security risk?
Yes. An extension with access to all sites can read and change every page a user opens, including signed-in email and business apps. Trusted extensions have received harmful updates after someone gained access to the developer's account, and copies of popular tools have collected data from their users.
How do I know if a Chrome extension is safe?
Check who publishes it, which permissions it asks for, whether it only reads or can also act on pages, where it sends data, and how often its owner or permissions change. An extension that needs little access, comes from a known company and has a clear data policy is lower risk. Nothing makes an extension safe for good, because updates can change it.
Is the Claude Chrome extension safe?
Claude in Chrome is published by Anthropic and has safety measures such as site permissions and confirmations before high-risk actions. It is still a browser agent that acts with the user's signed-in sessions, and its maker advises using it in a separate browser profile without banking, health or government accounts. Companies should decide centrally whether to allow it, for whom and on which sites.
What is extension risk?
Extension risk is the chance that a browser extension exposes data or takes harmful actions, combined with how bad that would be. It depends on the extension's permissions, its publisher, its update history, where it sends data and whether it can act on its own.
Is an ad blocker Chrome extension safe?
Ad blockers need to read and change pages to work, so they ask for broad access by design. A well-known blocker from an established publisher is usually acceptable, but the same checks apply: publisher, permissions, update history and a clear owner. On company devices, allow one approved blocker rather than whatever users find.
Are AI extensions riskier than other extensions?
Usually, yes. Most AI extensions send page content to an outside AI service, and agent extensions can also click, type and submit forms. Text on a web page can steer an agent, so the risk depends partly on every site the user visits.

Ready to act on it? Read how to block AI browser extensions, or check the official and third-party options in is the ChatGPT Chrome extension safe for work.

Claude and Claude in Chrome are trademarks of Anthropic, PBC. Perplexity is a trademark of Perplexity AI, Inc. ChatGPT is a trademark of OpenAI.

Monica and Merlin are trademarks of their respective owners. Google and Chrome are trademarks of Google LLC.

Jamf is a trademark of Jamf. Gartner is a trademark of Gartner, Inc.

Apple and Mac are trademarks of Apple Inc. Windows, Intune and Microsoft Edge are trademarks of Microsoft Corporation.

Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.