OpenAI's agent can browse authenticated sites and act on Scheduled Tasks triggers
OpenAI has removed one of the biggest roadblocks standing between ChatGPT and genuinely useful automation: the login screen.
Starting this week, ChatGPT Work can sign in to password-protected websites and carry out a task through to completion, instead of stalling out and handing control back to the user.
When ChatGPT Work hits an authentication wall, it now surfaces the site's login form directly rather than pausing the task entirely. Users type in their username, password, and any security code themselves, and OpenAI says that data never reaches the underlying model or gets used for training.
Existing password managers work within the flow too, and none of the credentials are stored afterwards.
Once a session is authenticated, it stays active for future tasks, so ChatGPT doesn't need to be walked through the same login twice.
Anyone who wants a clean slate can wipe browsing data for a single site or every connected site from ChatGPT's settings menu.
OpenAI is pitching this as a way to hand off genuinely tedious admin work. Listed use cases include setting up utilities for a new apartment, booking DMV or passport appointments, checking costs through an insurance portal, and pulling invoices from email to submit into accounting software.
Job recruiting tasks, like sourcing candidates and drafting outreach, are also on the list.
Even with login access granted, ChatGPT still stops to ask for confirmation before anything consequential, finalising a reservation or submitting a payment, for instance. The feature is rolling out now on web and mobile for Plus, Pro, and Business subscribers.
OpenAI paired the login update with a change to Scheduled Tasks: rather than firing only at set times, tasks can now trigger off real activity, including a new Gmail message, a Slack channel post, or GitHub pull request activity. Users can also share a scheduled task so someone else can adapt it to their own connected accounts.