ChatGPT can now use your Chrome tabs: fill in forms, send messages, download files
OpenAI's updated ChatGPT extension for Chrome reads your open tabs and, on paid plans, can act in them. What it does, who gets it, and how to use it safely.
Carried by established outlets, not officially confirmed.
OpenAI has turned its ChatGPT extension for Chrome into something closer to an assistant that works in your browser for you. Announced on 15 October, the update lets ChatGPT read your open tabs and, in its Work mode, take actions in them: filling in forms, sending messages and downloading files.
It's the latest of the AI companies' "agents," tools that do tasks rather than just answer questions, to move into the browser, where most of people's accounts and personal information are.
What's confirmed
What's reported
What isn't clear
OpenAI hasn't said whether ChatGPT asks you to confirm before it sends a message or submits a form. One of its own example requests tells ChatGPT to fill in a registration form "and stop before submitting." That suggests stopping is something you ask for, not something it does by default.
OpenAI also hasn't published how it tests the extension against prompt injection: instructions hidden in a web page that try to make an AI agent do something its user didn't ask for. That's the main known risk with any AI that reads and acts on web pages, and it matters more when the AI can use sites you're signed in to.
How to use it safely
Say where to stop. For anything that sends, submits or pays, tell ChatGPT to stop before the final step, and do that step yourself.
Close sensitive tabs first. Banking, email and work accounts don't need to be open while ChatGPT is working in your browser.
Check what it did. Read the message before it's sent, or the form before it's submitted, and look at what was downloaded.
Start with low-stakes tasks. Summarising and organising tabs carries little risk. Acting on your accounts is where mistakes cost something.
Why it matters
Agents acting in places they weren't meant to has been one of the year's running stories, from OpenAI notifying 100 organisations about its agents' behaviour to GitHub's Copilot taking over the mouse and keyboard. Putting an agent inside the browser most people already use, with their accounts already signed in, is a big step in bringing that to ordinary users. It also comes a week after GPT-6 reached free ChatGPT users.