Claude Duet: Anthropic details security and privacy

The company organizes the design around three principles: discretion with shared data, user control and credentials out of the agent's reach.

In this article you will see:

The protection layers of Claude Duet
How human approval and separate checking work
Why passwords and tokens stay out of the agent's reach
What the research preview still does not guarantee

The Claude Duet combines two layers and three principles

Anthropic describes two layers: safety training in the model and product safeguards. The article explains the new controls for tasks done in connected accounts and services.

Claude Duet can keep working even when the user closes the app. Anthropic identifies it as a research preview, not general availability.

Claude Duet asks for approval before sensitive actions

Before sending a message, paying or deleting something, the agent shows the action, the destination and the content, including the recipients. The user approves or rejects before execution.

A separate automated check evaluates whether the action matches the request. It runs outside the model that executes the task, so that hidden instructions on a page do not change this verification.

Diagram from Anthropic showing Claude Duet proposing an action, displaying an approval card and waiting for user review before execution — Imagem: Anthropic

Credit and source: Anthropic.

Anthropic keeps credentials out of Claude Duet

The company says the agent does not see passwords, browser cookies, login tokens or other credentials. This isolation follows the principle of separating secrets from the part of the system that does the work.

The article also describes a classifier that checks drafted messages before sending. For context on controls in agents, see how Claude Code fixed sandbox rules and how 1Password protects logins used by agents.

Claude Duet remains in research preview

Anthropic presents Claude Duet as an agent that can read information in connected services and execute tasks while the user is away. Sensitive actions, however, go through review before being completed.

The article describes controls designed to reduce risks, not a guarantee that every action will be safe. The company also says it continues working on product protections.

Frequently asked questions about Claude Duet

What is Claude Duet?

It is a research preview agent that can use connected services and keep working while the person is away, according to Anthropic.

Can Claude Duet send messages without authorization?

For sensitive actions described by Anthropic, no. The agent shows the action, destination and content and waits for the person's approval.

Can Claude Duet see passwords or login tokens?

No. Anthropic states that the agent has no access to passwords, browser cookies, login tokens or other credentials.

Is Claude Duet already available to everyone?

Anthropic's page describes it as a research preview and does not announce general availability.

Controls outside the model help keep the user in command

Claude Duet combines explicit approval, a separate check and credential isolation. For those building agents, the key point is to keep these barriers outside the model that executes the task.

From draft to finished product: split the work between specialist agents and track each step on bridgecode.