Dwelling
Newest reporting
Information
Anthropic detailed approval checks, information boundaries and credential isolation for Claude Duet, its cloud-based AI agent in analysis preview.
Anthropic stated Claude Duet, its cloud-based AI agent in analysis preview, is designed to behave on a person’s behalf whereas retaining approvals, delicate information checks and credential boundaries exterior the mannequin. In a submit revealed Oct. 23, 2026, the corporate stated the agent can learn inboxes, draft replies, e-book gadgets and run routines whereas a person is away, however is constructed to cease and ask permission for essential actions and to maintain passwords, cookies and sign-in tokens out of the mannequin’s attain.
Proof and context
The corporate stated Claude Duet is predicated on three working rules: person management, discretion over data shared by linked apps, and separation of credentials from the agent. In line with Anthropic, the system makes use of a separate classifier to evaluation messages earlier than they’re despatched and an automatic verify, constructed on its auto mode classifier from Claude Code and prolonged for Duet, to guage whether or not a process matches the person’s intent. The corporate stated that earlier than actions akin to sending delicate messages, paying for one thing or deleting gadgets, Claude Duet exhibits the vacation spot, content material and recipients, then waits for approval.
Anthropic stated some actions stay prohibited even with approval, together with securities buying and selling, banking actions akin to shifting cash between accounts and accessing another person’s account. The submit additionally stated sure repeated actions, akin to paying or sending cash, deleting an order or creating an account, usually are not coated by a “don’t ask me once more” rule. For purchases, Anthropic stated Claude Duet makes use of a one-time card from a pockets akin to Hyperlink from Stripe, restricted to the service provider and quantity, and that the person sees the merchandise and whole earlier than approving within the pockets app.
On privateness, Anthropic stated Claude Duet separates data into two tiers: “By no means share,” masking sign-in credentials, full card and checking account numbers and authorities ID numbers, and “Wants consent to share,” masking different non-public information akin to well being or monetary data. The corporate stated Claude Duet can rewrite a message with out delicate particulars whether it is blocked by both the sensitivity rule or consent requirement, however added that the safeguards can nonetheless miss one thing or flag innocent content material as a result of the product stays in analysis preview.
The submit additionally described visibility and management options. Anthropic stated customers can see what Claude Duet has accomplished in a process checklist, steer it mid-task, delete historical past or reset reminiscence, and disconnect apps so new entry stops instantly. It stated the agent inherits linked apps from the Claude account and makes use of reminiscence from prior Claude chats, with saved information wiped when the account is deleted.
Anthropic stated credentials are remoted from the mannequin and dealt with by separate machines. Passwords entered right into a type contained in the Claude app are encrypted with a single-use key on the cloud browser machine, then handed unopened to the browser. The corporate stated sign-in tokens for linked companies by no means enter the machine the place the agent works, and that the browser runs in its personal digital machine with fastened instructions akin to opening pages, clicking and typing. The submit stated the system can’t learn browser cookies, whereas web page textual content and downloads do come again as a result of the agent wants them to function.
Safety testing has been operating all through improvement, in line with Anthropic. The corporate stated its Safeguards group constructed evaluations from routine-shaped duties with harmless, injected and overbroad variants; crimson groups examined for leaks, unauthorized actions and malicious prompts; and an exterior penetration-testing agency reviewed the supply code, with a follow-on engagement on the cloud browser and sign-in type nonetheless beneath method. Anthropic stated Claude Duet is included in its public bug bounty program on HackerOne.
The product stays in a waitlisted U.S. analysis preview for Professional and Max subscribers, in line with background product data verified as of Oct. 23, 2026. Anthropic has additionally stated Claude Duet runs in standalone iOS and Android apps and on the net, and that it’s powered by Claude Opus 5.5. The corporate stated some capabilities and protections, together with approval cadence, sign-in flows and sensitivity checks, are nonetheless being refined because it gathers suggestions.
