ChatGPT Can Now Read and Control Your Chrome Tabs: Setup and Risks

OpenAI has turned its ChatGPT Chrome extension into something closer to a browser agent. On October 15, 2026 the company announced that ChatGPT can use the tabs you already have open: in Chat it can gather information across them, and in Work it can act in them. The Verge notes the context: "OpenAI's ChatGPT Atlas browser is gone," so the extension is now the main way ChatGPT reaches your browser. OpenAI's post on X said the goal is "bringing the power of computer use to anyone who uses Chrome," according to Unite.AI's summary of the official ChatGPT post.

This post explains what changed, how to set it up, what the permission prompts mean, and where the risks are.

TL;DR: Chat versus Work in Chrome

What changed in the October 15 update?

Three things, according to The Verge's report.

1. Tabs as context. When you tag Chrome in ChatGPT's web or desktop app, ChatGPT can view your open tabs. OpenAI's demo video shows someone asking it to compare three tabs of sofas they are considering.
2. Work can control tabs. In ChatGPT Work, the assistant can control your Chrome tabs for tasks such as "sending messages, filling out forms, or downloading files," and organize or group tabs. The Verge compares this to Google's Gemini-powered auto browse.
3. A full sidebar. The sidebar now carries "the full power of ChatGPT" with your complete conversation history in Chat or Work, so you can use it without leaving the page.

OpenAI's sample prompts show the intended uses: filling a registration form from an event brief and stopping before submitting, moving vendors from tabs into a comparison sheet with pricing and source links, comparing contractors and drafting an email to the best fit, and filling a dentist form while adding the appointment to work and personal calendars.

Why did Atlas go away?

Unite.AI, citing OpenAI release notes, reports that Atlas was deprecated on July 9, 2026 with a shutdown scheduled for August 9, and that users were pointed to the Chrome extension or sidebar. On August 31 the extension expanded to Edge, Brave, Opera and Vivaldi. Treat those dates as secondary reporting of OpenAI's notes; The Verge confirms only that Atlas is gone. The strategic read is that OpenAI chose to meet users inside the browsers they already use rather than ask them to adopt a new one. Unite.AI also notes the Chrome Web Store listing shows around 7 million users, a figure from the store page rather than OpenAI's announcement.

How do I set it up?

Per OpenAI documentation as summarized by Unite.AI:

1. Enable Computer Use in the ChatGPT desktop app's settings.
2. Install the ChatGPT extension from the Chrome Web Store.
3. Confirm that a Manage option appears.
4. In a ChatGPT Work or Codex chat, @-mention the browser.

Because many unrelated "ChatGPT sidebar" extensions exist in the store, check that the publisher is OpenAI before installing.

A sensible first run is a low-stakes, read-only task, such as asking ChatGPT to compare a few open tabs, before you let Work fill forms or send messages. That lets you see how the permission prompts behave on a site you do not care about, and you can decide per host whether "Allow once" is enough rather than reaching for "Allow for all sites."

What permissions does it ask for?

This is the part to read carefully. The Chrome install-time permissions listed include reading and changing data on all websites, browsing history across signed-in devices, bookmarks, downloads, notifications, tab groups, debugger access, and communication with native applications. That is a broad grant, typical for a browser agent but worth acknowledging.

At run time, the extension adds its own layer:

Allowlists and blocklists of domains are shared across supported browsers. By default the agent asks before interacting with each new website, and asks before downloading or uploading files. The store listing says it works in the background without taking over your active session.

What does OpenAI say about privacy?

The documentation, as summarized, says OpenAI does not keep a separate full record of browser actions; browser activity is stored only when it becomes part of ChatGPT context, such as page text, screenshots, tool calls, summaries or messages. Computer Use follows your Memories setting. The store listing says user data is not sold outside approved use cases. These are OpenAI's statements, not audited ones, and the record of agents acting without approval is a reason for caution: see our coverage of an agent sending an unapproved email and the analysis of AI agent privacy promises.

What are the real risks?

Prompt injection from web pages. A tab you did not write can contain text aimed at the agent. OpenAI's documentation advises treating page content as untrusted. If Work is allowed on a site where you are signed in, hostile text on any open tab is a possible route to your accounts.

Signed-in reach. OpenAI says ChatGPT can act on sites where you are already signed in, citing LinkedIn, Salesforce, Gmail and internal tools. That is the feature and the exposure.

Over-broad grants. "Allow for all sites" removes the main per-site check. Use per-site approval and a blocklist for banking, health and admin consoles.

Secrets in tabs. The documentation advises avoiding secrets in browser tasks unless required and reviewed.

For teams deploying browser or coding agents, AgentBeam, the agent security platform from the explainx.ai team, is built to stop agents before they take dangerous actions.

How does it compare with other browser agents?

Anthropic offers a similar extension; see our guide to Claude in Chrome and its real risks. Google has been adding Gemini to the browser itself, covered in Chrome's Gemini study tools and tab sync. OpenAI's differentiator is distribution: a single extension across five browsers plus the desktop app and its Work agent. Its ecosystem is also expanding; see ChatGPT Sites hosting MCP servers and plugins and the broader Intelligent UI changes in GPT-6.

A safe way to start

1. Begin with Chat, read-only, on a few research tabs.
2. Use Work only for tasks that stop before an irreversible action, as OpenAI's own form-filling example does ("stop before submitting").
3. Keep payment, email and admin tabs closed, or blocklisted, during agent tasks.
4. Review the Manage screen and revoke site permissions you no longer need.
5. Treat the first week as a test: log what it did against what you asked, and compare with what you would have done by hand.

What does this mean for how you work?

For people who live in tabs, the practical gain is removing copy-paste. Comparison shopping, vendor research and form entry are the tasks OpenAI chose to show because they involve many pages and a clear end state. The same properties make them good first tests: you can check the result quickly and the cost of a mistake is small if you stop before submitting.

For teams, the change shifts a decision from "should we allow an AI chat tool" to "should we allow an AI that can act in our logged-in SaaS tabs." That is a different policy question. IT administrators should check workspace settings for the extension, decide whether Allow for all sites can be restricted, and set a rule about which business apps may be open during agent tasks. Unite.AI's summary says the allowlist and blocklist are shared across supported browsers, which makes a baseline blocklist worth defining once.

For developers, the Codex mention in the setup steps is notable: the same browser connection is reachable from a coding-agent chat, so testing a web app in a real signed-in browser becomes part of the workflow. The store listing also advertises website testing and validation as a use.

What we do not know yet

OpenAI has not published error rates for Work in Chrome, nor a detailed account of its prompt-injection defenses in this release. Plan and region availability may vary, so verify in the extension itself. We will update this post as hands-on reports arrive, including any developer discussion on Hacker News.

Related reading

• Claude in Chrome: what the extension does and its risks
• Chrome adds Gemini study tools and tab sync
• ChatGPT agent sent an unapproved email
• AI agent privacy promises
• ChatGPT Sites, MCP servers and plugins
• GPT-6 and Intelligent UI in ChatGPT

Sources: The Verge, Unite.AI, and the ChatGPT account on X.

Details are as reported on October 15 and 16, 2026 and may change.