Source: ai-research/claude-blog-claude-duet-safety-privacy-security-2026-10-23.md (Anthropic, 2026-10-23), with ai-research/claude-help-get-started-with-claude-duet-2026-10-28.md and the launch post ai-research/claude-blog-introducing-claude-duet-full-2026-10-27.md.
Author: Anthropic · URL: https://claude.com/resources/articles/claude-duet-safety-privacy-and-security · Published: October 23, 2026
How Claude Duet is kept from acting against you while it reads your inbox and works unattended. Anthropic layers new Duet-specific controls on top of model safety training and the safeguards from Claude Code and Claude Managed Agents: approvals with a separate intent check, a sensitive-information classifier, and credentials that never reach the model. This page is the reference for what Duet will and won’t do on its own.
Three principles
- Your information, and what other people share in your connected apps, is handled with discretion, using a classifier that checks drafts before they go out.
- You stay in control of what it does on your behalf.
- Duet never sees your passwords, browser cookies, sign-in tokens or other credentials.
Approvals
- Before it sends a potentially sensitive message, pays, or deletes anything, Duet shows the action, the destination and the content, including every recipient, then waits.
- A separate automated check, built on the auto mode classifier from Claude Code and extended for Duet, judges whether the action matches your intent. It runs outside the model doing the work, “so instructions hidden in a page cannot affect the outcome.” Anthropic’s figure caption: “nothing the agent reads can answer the card.”
- Off-limits even with approval: trading securities; banking actions such as moving money between accounts; accessing someone else’s account.
- Never covered by “don’t ask me again”, however often you approve: paying or sending money, deleting an order, creating an account.
- Routines that run while you are away have the same safeguards.
- Anthropic calibrated how often Duet asks, because “approval fatigue” teaches people to tap through; the calibration will keep changing.
Sensitive information: two tiers
A classifier separate from the main model checks every message Duet is about to send, and files it uploads through the browser.
| Tier | Examples | Behaviour |
|---|---|---|
| Never share | Sign-in credentials, full card and bank account numbers, government ID numbers | Not shared with anyone, “even if you ask” |
| Needs consent | Health, finances, other private information | Shared only with your explicit approval; your own request counts (“send a note to my dentist about my tooth pain” generally needs no extra approval) |
- If blocked, Duet can rewrite the message without the sensitive part and bring it back for approval.
- Example Anthropic gives: a weekly team recap should include only what everyone on the team can already see, never your performance-review conversations.
- The classifier “will occasionally miss something or flag a harmless message”, which is why approvals stay on top of it.
Memory, history and disconnecting
- Each task has its own page in the task list; you can steer mid-task.
- Duet keeps conversation history and memory until you delete history or reset memory. Deleting your account wipes everything Duet stored.
- Apps connected in Claude carry over; switching one off in either place switches it off in both.
- Disconnecting stops new access only. What Duet already learned from that app stays until you reset its memory (from the app’s settings).
- Duet also uses your Claude account memory, so details from earlier Claude chats are visible to it.
Credentials and the browser
- Passwords. You type the password into a form inside the Claude app. It is encrypted with a single-use key that exists only on the cloud browser machine, passed unopened to the browser, and typed into the site. The model never sees it and can’t save or repeat it. 1Password sign-in is coming: you approve the request in 1Password and the credential goes straight to the browser.
- Phishing rule. Duet’s sign-in form appears only inside the app, during a sign-in you asked for. “It never arrives as a link, a text message, an email, or a website other than claude.ai.” Anything else asking for your password as Duet is not legitimate.
- Sign-in tokens. Following the Managed Agents design, tokens never enter the machine where the agent works; a placeholder sits there and a proxy outside inserts the real value as each request is made.
- Browser isolation. The browser runs in its own cloud VM, separate from the model and from the agent’s work computers. The agent drives it with a fixed command set (“open this page, click here, type this”); site code and cookies stay on the browser machine. Page text and downloads do come back, which is why the prompt-injection defences and the other checks exist.
- Identifies itself. Its browser sends a distinct User-Agent header and Anthropic publishes the IP addresses it browses from. Before launch, Anthropic offered several large sites the option to block Duet, “and some chose to.”
How it was tested
- Safeguards-team evaluations built from routine-shaped tasks, each in three variants: innocent, with injected instructions, and tempting the agent to overreach. These run alongside shadow traffic from real use.
- Red teams tested for leaks, unapproved actions and malicious prompts; a threat model covers the whole system.
- An external penetration-testing firm tested with source-code access; a follow-on engagement on the cloud browser and sign-in form was running at publication.
- Each new credential feature is red-teamed on an isolated rig before release.
- In scope of Anthropic’s bug bounty on HackerOne.
What is still changing
Anthropic lists three: how often Duet asks for approval, how you sign in, and how well the sensitive-information checks separate private from harmless.
Update 2026-11-01: sign-in and approval frequency in the field
- 1Password for agents announced. On The Artificial Intelligence Show Ep. 248 (recorded 2026-10-26;
raw/Ep._248_-_AI_Cybersecurity_Crisis_Dario_Fights_Back_on_Safety_OpenAI_s_Blind_Spots_Claude_Duet.md, https://www.youtube.com/watch?v=QY2qHT29ucc), Mike Kaput reported that 1Password announced “1Password for agents, which lets users approve an AI agent’s login while keeping passwords hidden from the agent.” The caption then reads “There are Muse integrations rolling out in Claude Duet and ChatGPT work integrations coming soon.” That most likely names integrations rolling out in Muse and Claude Duet, with ChatGPT Work to follow, but the wording is garbled. ^[ambiguous] Check Duet’s sign-in options before relying on it. - Approval frequency, one measurement. In Claire Vo’s nine-task benchmark, Duet asked for permission four times; Grok Bot, Muse and Dot asked once each. She called it “way too conservative.” See Claire Vo’s Personal Agent Benchmark.
Key Takeaways
- Approvals plus a separate, out-of-model intent check are the core control; a page Duet reads cannot approve its own action.
- Payments, sending money, deleting orders and creating accounts always need fresh approval; securities trading and moving money between accounts are never allowed.
- Credentials, full card numbers and government IDs are in a “never share” tier that your own request cannot override.
- Disconnecting an app does not erase what Duet learned from it; reset memory if that matters.
- Only enter a password into Duet’s in-app form during a sign-in you started.
Try It
- Before connecting a sensitive inbox, decide whether Duet should keep what it reads; if not, plan to reset memory after disconnecting.
- Test the approval card on a low-stakes send (an email to yourself) and read every recipient line before approving.
- If your agency or client sites block bots, check whether they block Duet’s published User-Agent and IPs before assigning browser tasks there.
Open Questions
- Anthropic did not publish the User-Agent string or the IP list location in the post; neither is captured here.
- Which large sites opted out of Duet visits is not named.
- How long memory and history are retained when you never delete them is not stated.
Related
- Claude Duet — the product overview.
- Payments with Link and Shop Pay — the one-time-card purchase flow.
- Auto Mode — the classifier Duet’s intent check extends.
- Claude Managed Agents — the credential-proxy design.
- Agent Guardrails — the wider pattern of approval gates and limits.
- Instruction Injection Layers — prompt-injection defence in depth.
- What Personal Agents Keep — the same disconnect-is-not-delete issue in another agent.