Claude Duet safety design: passwords and payments kept out of the model
Anthropic detailed Claude Duet's safety design: an out-of-model action checker, passwords never seen by the model, one-time merchant-locked cards, and non-dismissible confirmations for risky actions.
ANTHROPIC JUST SHOWED HOW CLAUDE DUET KEEPS YOUR PASSWORDS AND YOUR MONEY SAFE?! They published a full page on Claude Duet's safety. A separate check, built from Claude Code's auto mode classifier, tests whether each action matches what you asked for. It runs outside the model, so instructions hidden in a web page can't change its answer. More from the same page: · the model never sees your passwords, you type them into an encrypted form inside the app · purchases use a one-time Link card locked to one merchant and one amount · even with your approval it won't trade securities or move money between accounts · paying, deleting an order and creating an account can never be set to "don't ask again" Earlier this month I hid instructions for AI agents inside my own Claude Code test tasks: a fake IT email, a hidden div on a vendor page, a README demanding curl | sh, a comment inside a CSV. 135 runs. The agents followed 0. Would you let it pay for things on its own?
This post quotes another post: x.com/0xkense1/status/2113020432391520378