ChatGPT Work learns to log into sites without seeing passwords

OpenAI is launching today a website-login capability for ChatGPT Work, available in the browser and on mobile for Plus, Pro and Business subscribers. The model runs a virtual browser, feeds login details and navigates multi-step processes, the company says, without the password or username ever being exposed to it.

According to the announcement, the mechanism executes the login steps in an isolated environment: the user supplies the credentials once, and the agent uses them to carry out tasks such as setting up electricity and water accounts in a new apartment, booking a motor-vehicle office appointment or checking a refund status. OpenAI stresses that the model itself does not "see" the secrets; they are entered directly into form fields by the browser component, but the company has not released technical details on the architecture that separates the conversational context from the automation layer.

The rollout begins today (27 August 2026) in waves to all paid subscribers—Plus, Pro and Business—both in the web version and the mobile apps. Free-tier users are not included in the current round. OpenAI did not specify a timeline for extending the feature to additional tiers or to organizations outside the Business package.

An early user report points to immediate friction: the system encounters IP-based blocks and Cloudflare challenges that require verification “I’m not a robot”, leading to a repeated click loop. The issue is familiar from earlier browsing agents, as sites detect automated traffic patterns and block them at the network level, regardless of whether the credentials are valid. OpenAI has not responded yet to this limitation, and it is expected to limit usability on bot-protected interfaces during the initial launch.