What It Actually Does Now
The work runs inside a cloud browser that sits apart from your own device, so the agent is clicking around on OpenAI's infrastructure, not taking over your laptop. Once it can sign in, the list of errands gets a lot more practical. In its own announcement OpenAI points to setting up utilities for a new apartment, booking a DMV or passport appointment, and checking a reimbursement. WinCentral's writeup adds things like filing an insurance claim, restocking supplies, and filling out multi-step forms.
The reason this matters is simple. Most of the annoying tasks people actually want off their plate live behind a login. A shopping bot that can only browse public pages is a demo. One that can sign in and finish the job is a tool.
How It Handles Your Password
Credentials go in through a separate secure interface, not the chat window, and OpenAI says the password never becomes part of the normal conversation the model reads. So the pitch is that ChatGPT can be logged in as you without ever holding the key itself.
There are guardrails on top of that. Before anything with real consequences, a financial charge, a legal commitment, or a change to an account, it stops and asks you to approve. If a task hits a step it can't safely handle, it halts instead of guessing. You are still the one signing off on the parts that bite.
Who Gets It
Website sign-in is for paying users on Plus, Pro, and Business. Free accounts can still set up basic scheduled tasks, but they don't get the event-triggered automation or the login-only features. This is worth flagging because early summaries claimed free users got everything, and they don't.
The automation side got an upgrade too. ChatGPT Work can now be triggered by activity in Gmail, Slack, or GitHub, so a task can fire on its own when one of those services sees something new, rather than only running on a fixed schedule.
The Part Worth Watching
There's no password manager support yet, which is the first thing power users asked about, and a fair complaint. But the bigger question isn't a missing feature. It's trust. Handing an agent standing permission to log in as you and commit to things on your behalf is a real shift, and the approval prompts are the only thing standing between "it booked my appointment" and "it signed me up for something I didn't want." How well those checks hold up under everyday use is the thing to watch over the next few weeks.