ChatGPT Work cloud browser adds authenticated website sign-ins

OpenAI extended ChatGPT Work's cloud browser with authenticated website sign-ins, closing a major gap in agentic browser automation: the inability to complete tasks that live behind logins. Previously, agents could navigate public web pages but stalled the moment a workflow required authentication — which is most real enterprise and personal work (email, SaaS dashboards, banking, internal tools).

The mechanism is careful about security: a Work task encountering a login screen pauses, requests the user's credentials, and continues after authentication, with credentials going directly to the remote cloud browser rather than being exposed to or logged by the model. This human-in-the-loop-at-login design mirrors how enterprise RPA tools handle sensitive auth and is a deliberate response to the obvious risks of agents holding credentials.

The capability works across both web and mobile tasks, meaning an agent can complete end-to-end workflows that span authenticated services — a prerequisite for the 'ChatGPT Work does your real job' pitch OpenAI made when it launched the agent the day before.

This is a genuine continuation of the ChatGPT Work launch, adding a concrete new capability (authenticated sign-ins, dated August 26) rather than re-announcing the agent. It also directly parallels the week's agent-authentication theme: AWS's AgentCore giving agents payment credentials, and the broader push to let agents act with real-world authority. Competitively, authenticated browsing is exactly where Anthropic's Cowork, Google's Gemini-in-Chrome auto-browse, and Microsoft Copilot are all racing. The security model — credentials to the cloud browser, pause-for-input at login — will be scrutinized, since a compromised cloud browser holding user logins is a high-value target. Watch how enterprises weigh the productivity gains against the credential-custody risk.