What does the ChatGPT Chrome extension do?

Since 15 October 2026 it has two modes. Chat mode reads across your open tabs to answer questions, compare products or summarize pages, and you can point it at specific tabs with @chrome. Work mode uses computer use to operate your tabs: it can send messages, fill out forms, download files and organize tabs. Reports say it works with all ChatGPT plans, though availability can depend on rollout and workspace settings.

Is the ChatGPT Chrome extension safe?

It asks before acting on each new website, with options to allow once, allow for that site, allow everywhere or decline, and asks before using your history or moving files. The risks are the usual ones for agents: whatever it reads in your tabs goes to ChatGPT, a page can contain text written to steer it, and broad 'allow for all sites' permissions let it act wherever you are logged in.

Should companies allow the ChatGPT Chrome extension at work?

Decide it explicitly rather than by default. On internal tools that show customer, employee or financial data, Chat mode sends that page content to ChatGPT, so the plan your staff use matters: Business and Enterprise do not train on it by default, consumer plans may unless training is turned off. Work mode on those tools should be limited to specific sites and reversible actions.

How do I make my web app work well with AI agents like ChatGPT in Chrome?

Validate every action on the server, ask for confirmation before anything that costs money or changes data, make forms and buttons clearly labelled so an agent can read them, and consider exposing the main actions as WebMCP tools so agents call them directly instead of clicking. For assistants that work without a browser, offer an MCP server.

ChatGPT Chrome Extension Work Mode: What Businesses Should Know

Three browser tabs with a small robot arm moving between them filling a form, stopped by a glowing permission gate before the last tab

On 15 October 2026, OpenAI gave the ChatGPT Chrome extension a Work mode: it can now operate your open tabs, sending messages, filling out forms, downloading files and organizing tabs, through computer use. Chat mode reads across tabs to compare, summarize and answer. The extension already has millions of users, and reports say the update works on all ChatGPT plans. For a business this cuts both ways. Your staff will point it at internal tools full of customer data. Your customers will point it at your app and let it click through on their behalf. Below: what the extension does, what to watch inside your company, and how to make your own web app work with it instead of against it.

What changed on 15 October

From the coverage by AI Weekly and The Decoder:

Chat mode reads across open tabs: compare products by size, price or return policy, summarize articles, pull context from several pages. You can target specific tabs with @chrome.
Work mode operates tabs: it sends messages, fills out forms, downloads files and groups tabs. OpenAI's examples include registration forms, calendar events and business systems.
Permissions. It asks before each new website (allow once, allow for this site, allow for all sites, or decline), and before using your history or moving files.
Context. It takes over the computer-use features of OpenAI's standalone Atlas browser, which OpenAI shut down in August.

Inside your company: what to decide

Chat mode sends what is on the page to ChatGPT. On a CRM record, an HR system or a finance dashboard, that is customer, employee or financial data. Whether it is stored or used for training depends on the plan your staff are signed into: ChatGPT Business and Enterprise don't train on it by default, consumer plans may unless someone turned training off. We went through each plan in does ChatGPT store your data?

Work mode acts as the logged-in user. It can do anything that person can do in that tab, including on systems where one wrong click sends an email to a customer or approves a payment. "Allow for all sites" is the setting to avoid.

Pages can steer it. An agent that reads a page can be steered by text written to it, in a support ticket, an email or a review. That is the pattern behind this year's agent attacks, which we covered in AI agent security.

A short policy covers most of it: which plan staff must use, which internal tools Work mode may touch, and "allow per site, never for all sites".

On your own app: agents are coming as users

Your customers will soon let an agent fill in your forms and click your buttons while they watch. Four things decide whether that goes well:

Validate everything on the server. An agent submits whatever it understood. Your server decides what is allowed, exactly as it should for a human.
Confirm what costs money or changes data, with a clear summary of what will happen. A confirmation step helps the user stay in control of their agent.
Label forms and buttons properly. An agent reads the page. Unlabelled icons, fields with no names and multi-step flows hidden behind hover menus confuse agents the same way they confuse screen readers.
Don't block agents by reflex. Aggressive bot detection may now block a paying customer's assistant. Decide which actions you want agents to perform, and protect the rest.

Better than clicking: give agents tools

Clicking through a UI is the slowest, most fragile way for an agent to use your product. Two better options exist:

WebMCP, for agents in the browser: the page declares tools such as searchOrders or updateAddress, and the agent calls them directly, with your validation and confirmation built in. Shopify shipped exactly this in checkout in September. WebMCP explained.
An MCP server, for assistants that work without a browser: Claude, ChatGPT or your customers' own agents call your actions with the user's own permissions. How to build one.

Both start with the same design question: which actions should an agent be able to take in your product, with what permission, and with what confirmation?

ChatGPT Chrome Extension Work Mode: What It Does, and What It Means for Your Web App

On 15 October OpenAI gave its Chrome extension a Work mode that clicks, types and submits in your open tabs. Your staff will use it on internal tools, and your customers will use it on your app. Here is what it does, what to watch, and how to be ready on both sides.