Anthropic Details Claude Duet’s Safeguards as Browser Testing Continues
Sensitive actions pause for review, while credentials stay outside the model’s reach. Anthropic says testing of the cloud browser and sign-in form remains underway.

Anthropic outlined controls for Claude Duet on Oct. 23, describing approval checks and limits on sensitive tasks as the cloud-based assistant works across connected apps while users are away.
Duet is designed to pause before sending a message it flags as potentially sensitive, submitting a payment or deleting an item. The screen is meant to display the proposed action, its destination and its contents, including recipients, before asking the user to approve it. A separate automated review checks whether the action fits the user’s request. Some actions, including paying or sending money and deleting an order, cannot be added to the “don’t ask me again” setting, even after repeated approvals.
The system has two rules for private information. It is instructed never to share credentials, full card or bank-account numbers, or government identification numbers. Other private details, such as health or financial information, require the user’s consent. Duet is also barred from trading securities, moving money between accounts or accessing another person’s account, even when a user approves.
The Claude app encrypts passwords entered there using a key generated for one use. Browser cookies and sign-in tokens for connected apps are kept out of the model’s reach, and the cloud browser runs in its own virtual machine. Page text and downloaded files are returned to Duet so it can complete tasks.
Users can view and steer tasks while they are running. Disconnecting an app stops Duet from getting new access to it, but information the assistant has already learned remains until the user resets its memory. Anthropic warned that the protections “will occasionally miss something or flag a harmless message.”
Testing included evaluations using injected instructions and tasks that asked for too much, red-team exercises and a source-code penetration test by an external firm. A follow-up test of the cloud browser and sign-in form was still underway.
Anthropic opened the U.S. research preview on Oct. 21 to Pro and Max subscribers through a waitlist. Duet can keep working after a user closes the app. The company’s earlier announcement of the Claude Duet U.S. preview covered its launch; the Oct. 23 disclosure focused on the assistant’s safety and privacy controls.