To vet an AI browser extension, check which sites it can read (its host permissions), which extra powers it requests, such as clipboard or debugger access, and where your page text goes. The stakes are real: in December 2024 a phished developer account let attackers push a malicious update of the Cyberhaven extension to its users, and in December 2025 researchers at Koi Security reported that Urban VPN Proxy, with more than 6 million Chrome users, had been capturing conversations with ChatGPT, Claude, Gemini and other assistants since a July 9, 2025 update. Here is what the warnings mean and how to check an add-on in five minutes.
What extension permissions actually allow
Manifest V3 extensions declare access in a manifest file: permissions (named APIs), host_permissions (URL patterns for sites the extension may touch) and optional versions of both, requested later when a feature needs them.
Host permissions are the core of the risk. According to Chrome’s declare permissions guide, they let an extension inject scripts into matching pages, read sensitive tab details such as URL and title, watch and control network requests, read cookies through the cookies API, and make network requests from its background service worker. A pattern such as <all_urls> extends that to every site you open, including your bank and email.
scripting and cookies show no warning on their own because they only act where the extension already has host access; storage is silent too. activeTab is silent by design: it grants access to the current tab only after you click the extension, its menu item or shortcut, and ends when you move to another site. For an AI summarizer, activeTab plus scripting is the narrow option.
How AI extensions send page content to models
An AI sidebar or “summarize this page” button usually works in three steps. A content script inside the page reads the text or the field you are typing in, passes it to the extension’s background service worker, and the worker sends it to a model API, on the vendor’s server or a provider such as OpenAI or Anthropic.
So whatever the content script can read, from an open email to a medical portal, can leave your machine. Manifest V3 bans remotely hosted code, so an extension’s logic must ship inside the reviewed package, but remote data such as configuration files is still allowed, and collection can arrive in a routine update.
Comparison table: Chrome permission levels for AI extensions
Warnings are from Chrome’s permissions list and Web Store help.
| Permission | Warning shown | What it lets the extension do | Risk for an AI add-on |
|---|---|---|---|
activeTab |
None | Access the current tab after you click the extension | Low: you decide each time |
storage, sidePanel |
None | Save settings, show a side panel | Low |
scripting, cookies |
None on their own | Inject scripts, read cookies, only on sites it has host access to | Depends on host access |
| Host access to listed sites | Data on a list of websites | Read and modify pages on those domains | Medium: check the list |
tabs |
“Read your browsing history” | See the URL and title of every tab | Medium |
clipboardRead |
“Read data you copy and paste” | Read your clipboard | Medium to high |
| Host access to all sites | Data on all websites you visit | Read and change every page you open | High |
debugger, tabCapture, pageCapture, proxy |
“Read and change all your data on all websites” | Debug pages, capture tabs, route your traffic | Very high |
history |
“Read and change your browsing history on all signed-in devices” | Read and erase history across devices | High |
Worked example: reading a permission warning list
Take a hypothetical “AI page summarizer” whose install dialog shows these real Chrome warning strings: “Read and change all your data on all websites”, “Read data you copy and paste” and “Manage your downloads”.
The first warning means it can read every page you load, not just the one you summarize; Chrome also hides the tabs warning when all-sites access is requested, so tab tracking may be bundled in. Clipboard reading would expose passwords or codes you copy, and download management is unrelated to summarizing. Web Store policy asks developers for the narrowest permissions a feature needs, and a summarizer using only activeTab and scripting would install with no warnings. This one fails the test: pick an alternative, or limit its site access after install.
What real incidents teach
Cyberhaven shows a trusted extension can turn bad overnight. According to Obsidian Security and SecurityWeek, an employee was phished by an email posing as Chrome Web Store support and authorized a malicious OAuth app, “Privacy Policy Extension”. The attacker uploaded version 24.10.4, which was live for more than 24 hours across December 25 and 26, 2024, until clean version 24.10.5 replaced it. It stole Facebook advertising access tokens, account details and cookies. SecurityWeek reported that Internxt VPN, VPNCity, Uvoice and ParrotTalks were hit in the same wave. Chrome only pauses an extension when an update adds a new warning, so updates that reuse existing permissions install silently.
Urban VPN shows a “privacy” product can harvest AI chats. The Register reported that it overrode browser network functions on AI sites, collected by default with no toggle and sent data to an affiliate of data broker BiScience, while holding a Featured badge. Google’s Limited Use policy forbids transferring user data to data brokers, and Mozilla requires opt-in before personal data is sent, but policy is not a guarantee. For a coding-tool parallel, see our report on the backdoored Claude Code extension on Cursor’s marketplace.
Step-by-step checklist before you install
- Read the install warnings and compare them with the one feature you want.
- Prefer click-to-run access. Extensions using
activeTabor optional permissions ask only when needed. - Open the listing’s privacy tab to see which data is collected and whether Limited Use compliance is certified.
- Find the model provider and whether your text is kept or used for training; our guide to what happens to what you tell an AI chatbot covers what to look for.
- Check the publisher. A known company beats an anonymous developer, though Cyberhaven shows it is no immunity.
- Restrict site access after install to “On click” or a short list of sites, keeping banking, health and work systems off it.
What to watch next
- Whether Google tightens Featured badge and Limited Use enforcement for AI-chat readers.
- Mozilla’s built-in data consent prompt for Firefox add-ons, which requires manifests to declare data collection.
- More phishing campaigns aimed at extension developers’ store accounts.
Key takeaways
- Host permissions decide what an extension can read; all-sites access means every page.
activeTabgives AI tools page access only when you click, with no install warning.- Updates can turn a clean extension malicious without any new permission prompt.
- Store badges and policies such as Limited Use reduce risk but did not stop the Urban VPN case.
- After installing, restrict site access to “On click” or specific sites.
Frequently asked questions
Is “Read and change all your data on all websites” always dangerous?
It is not proof of bad intent, but the extension could read every page you load. For an AI tool that works on the current page only, that is more than it needs.
Do AI browser extensions send my page content to a server?
Most do, because the model usually runs on remote servers, not in your browser. The privacy policy should say which service receives your text and how long it is kept.
Can I limit an extension’s access without uninstalling it?
Yes. In Chrome, open Manage extensions, choose Details and set site access to “On click” or “On specific sites”.




