From the source
Anthropic says Claude Duet, its research-preview agent that works from a cloud computer, handles user information under three principles: approval before important actions, discretion about what it shares, and isolation of credentials from the model.
- Before sending a potentially sensitive message, paying for something or deleting anything, Claude Duet shows the action, destination and content, including every recipient, then waits for approval.
- A separate automated check built on the auto mode classifier from Claude Code runs outside the model doing the work, so instructions hidden in a page cannot affect the outcome.
- Some actions stay off-limits even with approval, including trading securities, moving money between accounts and accessing someone else's account, and paying or sending money is never covered by a "don't ask me again" rule.
- A new safety classifier sorts information into two tiers: sign-in credentials, full card and bank account numbers and government ID numbers are never shared, while health or finances need explicit consent.
- Anthropic says the safeguards will occasionally miss something or flag a harmless message, and that it expects to keep iterating on the classifier.
