ChatGPT in Chrome: What the New Tab Access Can Do, and Its Limits

ChatGPT can now see what you have open in your browser. OpenAI has rebuilt its Google Chrome extension so the chatbot can read your open tabs and, if you ask, work inside them: filling out forms, sending messages, downloading files, organizing tabs. That brings part of what OpenAI tried with its own AI browser, Atlas, into the browser most people already use.

Key takeaways

• The ChatGPT extension for Chrome can now draw on your open tabs. You target the browser directly with an @-mention in the chat.
• In Chat mode, ChatGPT only reads and gathers information across several tabs. In Work mode, it controls the tabs itself.
• OpenAI says the updated extension is available on all ChatGPT plans. It has not said whether tab control works the same way everywhere.
• By default, ChatGPT asks for permission before acting on each new website. OpenAI itself classifies page content as untrusted.

What the extension can do now

OpenAI introduced the changes on October 15 in a thread on X. At its core, the extension now has two ways of working. In Chat mode, ChatGPT looks across your open tabs and pulls together what it needs for an answer. If you have five stores with sofas open, you can have it line the models up by dimensions, material, delivery time, and return policy without clicking through each page. OpenAI uses exactly that example, along with comparing pricing pages from different vendors and summarizing long articles in plain language.

Work mode goes a step further. Here ChatGPT uses computer use, a model’s ability to operate a screen and inputs the way a person would, to act inside the tabs. OpenAI lists sending messages, filling out forms, downloading files, and grouping tabs. As further examples, the company mentions copying vendor details from several pages into a spreadsheet and adding appointments to a calendar. There is also a new sidebar that shows the full chat history next to the page you are viewing. It resembles the Gemini panel Google has built directly into Chrome.

Strategically, this is a change of course that was a while coming. In July, OpenAI announced it would shut down its Atlas AI browser after nine months. Instead of persuading people to switch browsers, OpenAI is now betting on an extension for the market leader, whose maker, Google, runs a competing assistant in the very same window with Gemini.

How to set it up

OpenAI’s help page describes setup through the ChatGPT desktop app. Under Settings and Computer Use, you pick your browser and install the extension from the relevant store using the button shown there. Once a “Manage” option appears in settings, the connection is in place. You then start a Work or Codex chat and mention the browser with the @ sign. It matters that you use the browser profile in which the extension is installed.

According to the documentation, the extension supports not only Chrome but also Edge, Brave, Opera, and Vivaldi. The sidebar, which OpenAI calls Side Chat, is not available in Opera. On a Mac, it opens with Cmd+Shift+Period. If you want ChatGPT to upload files, you also have to allow the extension access to file URLs in Chrome. OpenAI notes that availability can vary by rollout and workspace settings, so business users should check with their IT department first.

Who benefits? Mostly people who regularly compare, research, or carry data from one page to the next: checking offers, assembling travel details, sorting research results into a spreadsheet. Chat mode is the low-risk way in because it only reads. Early on, Work mode requests are best phrased with clear limits, the way OpenAI demonstrates it: fill out a form and stop before submitting.

Permissions and limits

A browser agent is only as trustworthy as the pages it reads. That is why OpenAI added a per-site prompt: before ChatGPT interacts with a new domain, it asks. The choices are Allow once, Allow for this site, Allow for all sites, and Decline. OpenAI itself flags allowing all sites as higher risk. There is no standing permission for browsing history at all; ChatGPT asks every time. Allowed and blocked sites can be managed in settings.

The documentation is notably candid on another point. It explicitly treats page content, highlighted text, and video transcripts as untrusted. Misleading or malicious content could raise the risk that data gets copied to places it was never meant to go. This is the problem of prompt injection: hidden instructions on a web page that an AI agent mistakes for a request from the user. OpenAI has not published figures on how well the extension resists such attacks. For comparison, when Anthropic launched Claude for Chrome in August 2025, it reported an attack success rate of 23.6 percent in internal tests without safeguards and 11.2 percent with them.

The scope of permissions is worth a look, too. The extension requests broad browser permissions, including debugger access, reading and changing data on all websites, history, bookmarks, and downloads. OpenAI says it does not keep a separate, complete log of all browser actions. What gets stored is whatever becomes part of the chat: text that was read, screenshots, tool calls, or summaries. OpenAI advises users not to route passwords and other sensitive data through browser tasks.

Outlook: the browser becomes the agent’s workplace

This update shifts the question of where AI agents do their work. Not in a specialty browser that few people install, but in the environment where people are already signed in, with their accounts, tabs, and bookmarks. That makes the extension useful right away, because the agent finds exactly what the user has in front of them. It also makes it more sensitive than an isolated test browser, because every approved site is a real, logged-in session. The sensible division of labor almost suggests itself: be generous with reading and comparing, and approve actions site by site. The interesting part is how Google responds. It controls the browser OpenAI is now working in, and with Gemini in Chrome it has the more direct line to the user.