Claude Duet has a short, strict set of approval rules. Before it sends, pays, signs or deletes anything, it shows you the exact action, every recipient and the content, then waits. Paying, deleting an order and creating an account always ask and cannot be set to "don't ask me again". Trading securities, moving money between accounts and accessing someone else's account are refused even if you approve. Most of your control comes from which apps you connect and how carefully you read each request. This guide is based on Anthropic's published docs as of November 2026; Duet is a US-only research preview that launched on 21 October 2026.
How do Duet's approvals work?
Anthropic's safety article (23 October 2026) describes two layers.
First, an approval screen. For consequential actions Duet lays out what it is about to do, who will receive it and the full content, and waits for you.
Second, a separate action check that runs outside the model. It is based on the classifier used in Claude Code's auto mode. Anthropic says "Instructions hidden in a page cannot affect the outcome" of that check, because the check does not take instructions from page content.
- Duet plans an action
- Separate action check
- Approval screen with recipients and content
- You approve or decline
Anthropic is also clear that this is not a guarantee: it says no defence is perfect, which is one reason consequential actions need your approval. Its browser-use research reported a 1.4% attack success rate for Claude Opus 4.5 with safeguards, and called even 1% "meaningful risk".
What always asks, and what never runs?
| Category | Actions | Can you change it? |
|---|---|---|
| Asks before acting | Sending messages, signing documents, other deletions | You approve each time; some can be set to "don't ask me again" |
| Always asks | Paying or sending money, deleting an order, creating an account | No |
| Never runs | Trading securities, moving money between accounts, accessing someone else's account | No, even with approval |
Based on Anthropic's safety article and help centre, as of November 2026. Payments sit only in the "always asks" row: they can never be set to "don't ask me again".
What about "don't ask me again"?
Some lower-stakes approvals can be put on "don't ask me again". Be selective. A standing approval removes your chance to spot a wrong recipient or an injected instruction later. Read why "Always allow" is risky before using it, and never use it for anything that sends to people outside your usual contacts.
How should I read a Duet approval?
- Check the actionIs it the thing you asked for, on the site you expected?
- Check every recipientLook for unfamiliar addresses or extra names.
- Read the contentLook for private details the task did not need.
- Check the amountFor purchases, confirm merchant and total.
- Decline if unsureYou can ask Duet to explain or change the plan.
Approvals appear in the main thread and in the Tasks tab, and they come back if you do not answer. Only you can mark a task complete, so Duet cannot quietly declare victory on something it did not finish.
How do purchases work?
Duet always asks before paying. It uses a one-time card limited to one merchant and one amount, through Link or Shop Pay, so a mistaken or manipulated purchase cannot be reused elsewhere. If something you did not intend is charged, contact your card issuer and see our guide to an unauthorized AI agent purchase.
Which settings matter most?
Connectors
Connectors carry over from your Claude account. Anthropic says switching an app off in Claude also switches it off in Duet. Before using Duet, review your Claude connectors and switch off anything Duet does not need. Microsoft 365 was not yet available at launch. See least privilege for AI agents.
Routines
Routines, which run on a schedule, use the same safeguards as tasks you start. Approval requests from a routine still wait for you.
Sign-in
Duet never sees passwords, cookies or sign-in tokens; a proxy swaps tokens in. The sign-in form appears only in the app and "never arrives as a link, a text message, an email, or a website other than claude.ai". Any message asking you to sign in for Duet elsewhere is a scam. See agent takeover, passwords and 2FA.
Sensitive information
Anthropic says a classifier watches for sensitive information. It is a safety net, not a reason to share more than a task needs.
What about data and memory?
- Retention. Anthropic says Duet keeps data until you delete it or reset Duet.
- Disconnecting. Switching off a connector stops new access but Duet keeps what it already learned. Reset or delete to clear it. See deleting AI agent memory.
- Training. Anthropic has no Duet-specific training statement, but its consumer policy trains only if the Model Improvement setting is on, and that most likely applies to Duet. Applying it to Duet is an inference, so check the setting yourself.
Who can use Duet?
Duet is for adults (18+) in the US, as a research preview. Anthropic runs a HackerOne bug bounty for security reports.
Pros
- Exact recipients and content shown before every consequential action
- Hard refusals for trading, transfers and other people's accounts
- Action check runs outside the model
Cons
- No live view or takeover of the browser yet
- "Don't ask me again" exists for some actions and is easy to overuse
- Disconnecting a connector does not erase what Duet learned
What should I do if Duet does something unexpected?
Decline any pending approval, check the Tasks tab for what else is running, and switch off the relevant connector in Claude. If a charge appeared that you did not intend, contact your card issuer. Report the problem to Anthropic support and keep the task history. Our checklist for when an agent did something you didn't ask covers the rest.
Bottom line
Duet's defaults are among the strictest of the major agents. Keep them that way: limit connectors in Claude, use "don't ask me again" sparingly, read every recipient, and reset Duet if you want it to forget. For an overall verdict, see our sister site's Claude Duet review, and for how Duet compares, our safety features comparison.



