ChatGPT can now click, type and send inside the browser tabs where you are already logged in: your inbox, your CRM, your supplier portals. That is genuinely useful for the dull web chores that eat a small business owner's week. It also means handing an AI the keys to your signed-in accounts. The setup matters more than the features.
What Changed
On October 16, The Decoder reported that ChatGPT's Chrome extension had become a full browser assistant. In Chat mode, it pulls information from several open tabs at once; OpenAI's examples include comparing products by size, material, shipping time and return rules. In Work mode, it controls your tabs through Computer Use (OpenAI's term for AI that operates software by clicking and typing) to send messages, fill out forms, download files and group tabs. The Decoder says it works with all ChatGPT plans.
OpenAI's help page spells out the part that matters most: "ChatGPT can read or act on sites where you’re already signed in, such as LinkedIn, Salesforce, Gmail, or internal tools." You set it up from the ChatGPT desktop app, and it also works in Edge, Brave, Opera and Vivaldi.
Why Signed-In Tabs Raise the Stakes
An agent working in your logged-in tabs can do anything you could do there. The biggest risk is prompt injection: instructions hidden in a web page or email that try to steer the AI into something you never asked for. OpenAI's own guidance: "Treat page content as untrusted context." In its note on browser history, it warns: "Malicious or misleading page content can increase the risk that ChatGPT copies this data somewhere unintended."
Plain mistakes matter too. Last month OpenAI cancelled a planned model, GPT-6.1 Astra. Engadget, relaying The Wall Street Journal, reported that in testing it took actions "without asking for permission, such as using external tools and services." If the labs worry about that, so should you.
Never Let It Touch These
- Banking, payments and payroll. One wrong click moves real money.
- Admin panels for your website, domain, email or office suite. A mistake there can lock everyone out.
- Password managers and sensitive customer records. OpenAI's help page says: "Avoid sending secrets or highly sensitive data through browser tasks unless they’re required and you are present to review each prompt."
- Your browser history. OpenAI labels sharing it "Elevated Risk." Say no when asked unless the task truly needs it.
The Four Permission Choices
"By default, ChatGPT asks before it interacts with each new website," OpenAI says. When it asks, you get four options:
| Choice | What it does | Use it for |
|---|---|---|
| Allow once | Uses the site this one time. | Anything you want help with but would rather watch. |
| Allow for this site | Uses the site again without asking. | Low-risk sites, like supplier catalogs or public research. |
| Allow for all sites | Stops asking anywhere. OpenAI marks it "Elevated Risk." | Nothing, in a business setting. |
| Decline | Stops it using the site. | Banking, payroll and admin panels. |
Choices from OpenAI's browser extension help page. The "Use it for" column is our recommendation.
A Safe Setup in Five Steps
- Create a separate browser profile for AI work, signed in only to the sites you want ChatGPT to use. The extension works in the profile where you install it, so your everyday profile, with your bank logins, stays separate.
- Block your sensitive sites first. In the desktop app, open Settings, then Computer Use, then Manage, and add your bank, payroll and admin domains to the blocklist.
- Keep the default prompts on. Never choose "Allow for all sites."
- Watch the first ten tasks and check anything before it is sent or submitted.
- Write the rules down: which sites are allowed and who may use it. Our one-afternoon AI policy has a section for approved tools.
OpenAI notes that availability can depend on rollout and workspace settings. On a Business or Enterprise plan, check with whoever manages your account.
Good First Jobs
Start where a mistake is cheap and easy to spot: comparing supplier prices across tabs, pulling details from a few pages into one summary, filling a vendor form you review before submitting, or sorting 30 open tabs into groups. For how other browser agents compare, see our browser agents field guide. And if a chore repeats every week, a proper workflow automation is often more reliable than an agent clicking through it each time.
Sources
- The Decoder, "ChatGPT becomes a full Chrome browser assistant that reads and acts on your open tabs" (October 16, 2026): Chat and Work modes, examples, plans. the-decoder.com
- OpenAI, "Browser extension" (ChatGPT help page): signed-in sites, setup, permission choices, warnings. learn.chatgpt.com
- Internet Archive, snapshot of OpenAI's browser extension help page (October 17, 2026): the same guidance as published at the time. web.archive.org
- Engadget, "OpenAI reportedly cancels GPT-6.1 Astra's release over deceptive behavior" (September 29, 2026): actions taken without permission. engadget.com
Want AI handling your web chores safely?
Tell us which portal and form-filling jobs eat your week. We will set up the right mix of browser agent and automation, with guardrails your team can follow.