ChatGPT Work can now log into websites and complete tasks autonomously

OpenAI on Monday rolled out authentication support for its ChatGPT Work cloud browser, enabling AI agents to log into websites and complete tasks that previously required users to hand-hold the process through every password-protected page.

How It Works

The update, announced August 25, addresses a limitation OpenAI's own documentation had previously spelled out: the cloud browser would not accept credentials or sign in anywhere. Now, when the agent encounters a login page, it pauses and presents a secure form where users type their username, password, and any two-factor authentication code themselves. OpenAI states that credentials "go directly to the remote browser" and are never seen by the model, stored, or used for training.neowin+1

Password managers are supported, and a separate review model checks whether the destination page looks like phishing before the form appears. Once authenticated, the session persists for future tasks until it expires, eliminating the need to log in repeatedly.notebookcheck+1

Use Cases and Guardrails

OpenAI highlighted scenarios now possible with authentication in place: booking DMV or medical appointments, checking insurance costs, submitting reimbursement paperwork, finding job candidates, and managing invoices across platforms. Even with a live session, ChatGPT still asks for explicit user confirmation before consequential actions such as completing reservations or making payments.neowin

The feature is available to ChatGPT Plus, Pro, and Business subscribers on web and mobile. Users can clear browser data for individual sites or all sites under Settings to sign out.notebookcheck+1

Security Trade-offs and Additional Updates

The convenience comes with a residual risk OpenAI itself acknowledges. The cloud browser runs on a remote machine that keeps working after a user's device is off, and the session cookie sitting there is enough to access an account without a password or second factor. OpenAI tests for prompt injection and unintended actions but notes those "safeguards do not eliminate every risk".notebookcheck

Alongside the login feature, OpenAI announced that its Scheduled Tasks can now respond to events — including new Gmail messages, Slack messages, and GitHub pull request activity — rather than only running at set times. Scheduled Tasks is also expanding to free-tier users, limited to three active tasks with no webhook triggers and a maximum of one run per day.neowin