Claude Duet won’t share your card number even if you ask

Anthropic on October 24 explained how Claude Duet, its research-preview agent that keeps working in the cloud after you close the app, is kept away from your credentials and made to stop before risky actions, in a post on Claude’s site. Steps such as paying or sending money, deleting an order or creating an account come back for approval every time, however often they were approved before. Disconnecting an app cuts off Duet’s new access at once, but what the agent already learned from that app stays in its memory until the memory is reset.
Duet’s sign-in form appears only inside the Claude app, during a sign-in you asked for, and never arrives as a link, a text message, an email or a website other than claude.ai; anything that looks like Duet asking for a password elsewhere is not legitimate. Since Duet holds no credentials for you, it can’t be phished for a password, and no hidden instruction can make it reveal a secret it doesn’t have. The sign-in form, though, is itself a target, because anything that asks for a password can be imitated.
What always needs a yes
Before Duet sends a potentially sensitive message, pays for something or deletes anything, it shows the action, the destination and the content, every recipient included, and waits for approval. Trading securities, banking actions such as moving money between accounts, and accessing someone else’s account stay off-limits even when approved. Routines that run while you are away carry the same safeguards.
A separate automated check runs alongside each action and assesses whether it matches your intent. It is built on the auto mode classifier from Claude Code and extended for Duet, and like the approval card it sits outside the model doing the work, so instructions hidden in a page can’t affect the outcome and nothing the agent reads can answer the card. In Claude Code, the auto mode classifier that vets each action sees only the user’s messages and the agent’s tool calls, with Claude’s own messages and tool outputs stripped out, which makes it “reasoning-blind by design,” Anthropic wrote in its post on auto mode.
Anthropic says it spent a lot of time calibrating how often Duet asks, because too many requests “teach people to tap through them, sometimes without adequate consideration,” and that it will keep adjusting. The same auto mode post put a number on the habit: Claude Code users accept 93% of manual permission prompts.
For purchases, Duet pays with a one-time card from a wallet like Link from Stripe, limited to that merchant and amount. You see the item and the total and approve the purchase in your Link app, and your real card number reaches neither Duet nor the merchant.
Memory and connected apps
Each task has its own page in Duet’s task list and can be steered midway. Duet keeps conversation history and memory to carry context from one task to the next until the history is deleted or reset, and deleting the account wipes everything it stored. It also uses the memory of your Claude account, details Claude remembered from earlier chats included.
Apps connected to Claude carry over to Duet, and switching one off in either place switches it off in both. Clearing what Duet learned from a disconnected app takes a memory reset, available anytime from the app’s settings.
What Duet won’t send
A new classifier, separate from the main model, checks every message Duet drafts for sensitive information before it goes out. In Anthropic’s example, a weekly recap for your team should leave out private details such as your performance review conversations and include only what everyone on the team already has access to.
Sign-in credentials, full card and bank account numbers and government ID numbers get the strictest tier: Duet is instructed never to share them with anyone, even if you ask. Health, finances and other private information go out only with explicit approval, and your own request counts as consent, so “send a note to my dentist about my tooth pain” generally won’t need another one. When a message is blocked, Duet can rewrite it without the sensitive information and bring it back for approval, and files it uploads through the browser get the same check.
These safeguards will occasionally miss something or flag a harmless message, Anthropic says, which is why they work alongside the approvals. How well the checks tell what’s private from what’s harmless is still changing in the preview, along with how often Duet asks for approval and how you sign in.
Passwords, tokens, the browser
Duet never sees your passwords, browser cookies or the sign-in tokens for your connected apps, and it can’t save them to its memory. They pass only through separate machines the model can’t access, where the approvals and safety checks run too. A password goes into a form inside the Claude app, which encrypts it with a single-use key that exists only on Anthropic’s secure cloud browser machine; the form hands it over unopened, and the browser types it into the site. Sign-in through 1Password is coming soon: Duet will request a credential, you approve the request in 1Password, and the credential reaches the browser without passing through the model.
Sign-in tokens follow the Claude Managed Agents design: a placeholder sits on the machine where the agent works, and a proxy outside it puts in the real value only as a request is made. That design was Anthropic’s fix for the first version of Managed Agents, which ran Claude’s generated code in the same container as the credentials, so a prompt injection “only had to convince Claude to read its own environment.”
The browser runs in its own virtual machine in the cloud, apart from the model and from the computers the agent works on, and the agent drives it only through fixed commands: open this page, click here, type this. Site code and cookies stay on that machine, and none of the agent’s commands can read cookies out of it. The page’s text still comes back, because the agent has to read it, which is why the other checks exist and why Duet also uses the prompt-injection safeguards from Anthropic’s browser-use research. In that research, published in November 2025, Anthropic wrote that no browser agent is immune to prompt injection and that “a 1% attack success rate—while a significant improvement—still represents meaningful risk.”
Requests from Duet’s browser carry a distinct User-Agent header, and Anthropic publishes the public IP addresses it browses from, so sites can recognize the agent and decide how to treat it. Before launch, Anthropic offered several large websites the option to keep Duet out, and some took it.
Who has tested it
Anthropic’s Safeguards team built evaluations from routine-shaped tasks, each in three variants: an innocent one, one with injected instructions and one that tempts the agent to do more than it was asked; they run alongside shadow traffic from real use. Red teams checked whether Duet would leak information, act without approval or follow malicious prompts, and every new credential feature is red-teamed on an isolated rig before it ships to anyone.
An external penetration-testing firm tested Duet with access to Anthropic’s source code, and a follow-on engagement covering the cloud browser and the sign-in form is running now. Security vulnerabilities go to Anthropic’s public bug bounty program on HackerOne, which covers Duet, rewards valid reports and offers safe harbor for good-faith research.
Up next
ClaudeClaude Codeanalysis
Claude Code won’t type passwords, even in a test env
Claude Code refused to type a password into a dev’s own test environment, per a GitHub bug report, and explicit authorization didn’t budge the system prompt rule. When your agent hits a password field, do you type it in or script around it?
Related stories
- Testing Catalog: Claude iOS app gets “Saved logins” list
- Claude Duet works while you’re away and asks before acting
- Claude Duet users get 3 invites each, US only
- Claude’s auto mode skips connector “Always ask” settings
- Anthropic’s test agents filled out US visa applications
- Claude desktop teardown: MCP servers run outside the VM
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

