Claude Duet Adds Safety Safeguards Before Users Trust It With Tasks

At a glance

Claude Duet is a research-stage agent that completes tasks by using connected apps on your behalf. It keeps running on a cloud computer even after the user closes the app. Anthropic said it built this agent with privacy and security as basic requirements.

Key points

1. 1

   Important actions, such as sending sensitive messages, making payments, or deleting things, show the content and recipient and wait for user approval before running.

2. 2

   Each action goes through a separate automatic check that confirms it matches the user's intent, and this check runs outside the model doing the work.

3. 3

   Some actions are not allowed even with approval, such as stock trading or transfers between accounts.

4. 4

   Payments use a single-use card limited to a specific merchant and amount, so the real card number is not shared with the agent or the merchant.

Why it matters

If AI reads your email and makes payments for you, safeguards are essential to stop your information from leaking to the wrong places.