ChatGPT can now log into your online accounts without ever seeing your password
Key Points
- ChatGPT Work's cloud browser can now sign in to websites on web and mobile
- Users type their own credentials into a secure form; OpenAI says the model never sees them
- The signed-in session persists on OpenAI's servers until it expires
- Live from 25 August for Plus, Pro and Business subscribers in supported regions
- Permissions and site-by-site blocking sit in Settings > Cloud browser
OpenAI has switched on website sign-ins in ChatGPT Work, allowing the AI tool to access work which was previously protected by a login.
The change went live on Tuesday (25 August) for Plus, Pro and Business subscribers on web and mobile. It applies to ChatGPT’s cloud browser, which runs on OpenAI’s own machines and clicks through tasks on a user’s behalf.
Until this week the browser was only capable of handling public pages. This means it could not accept credentials, use a password manager or sign in anywhere.
As of this update, when the agent reaches a login screen it stops and shows a secure form, and the user enters the username, password and any two-factor code themselves.
OpenAI said the model never sees those credentials and does not store them.
The authentication will persist for future tasks until it expires, so users do not have to sign in again each time. This is done by leaving the cookie carrying the login sitting in a browser on OpenAI’s servers.
OpenAI said that this will open up several previously undoable tasks, including:
- Setting up utilities at a new address
- Booking a passport or doctor’s appointment
- Checking what an insurer will reimburse
- Preparing car registration renewal paperwork.
- Returns a nd bookings
- Arranging a package pickup
- Cancelling tickets after a trip changes
- Booking a vet appointment
- Submitting reimbursement paperwork for medical treatment.
To access the feature, users need to click on Settings > Cloud browser. The default settings asks before ChatGPT visits any new website, while an auto-approve option lets ChatGPT check the URL itself and only pause on something that looks unsafe. A third setting allows every website, which OpenAI itself does not recommend.
Individual sites can be allowed or blocked separately, and those choices override the default. Parents and guardians can switch the cloud browser off for linked teen accounts through parental controls.
Site access and action approval remain separate. ChatGPT will still ask for confirmation before anything with a financial, legal or account consequence, including completing a booking or making a payment.
OpenAI said that inital availability will be across paid plans in supported regions, excluding Free and Go tiers.
OpenAI said it has tested the system against prompt injection, phishing and unintended actions, but cautions that those safeguards do not remove every risk.
