AgentDuetUse casesTemplatesPricingFAQAdd to Chrome — try free
Guide · 6 min read

Is it safe to let an AI click for you?

What can go wrong with browser agents, and the safeguards Claude Duet uses: approvals, no passwords or card numbers, visible steps and a stop button.

The real risks

An agent works with your browser session, so it can do what you can do on a site. The main risks are ordinary ones: picking the wrong option, submitting a form too early, or following a misleading instruction written on a web page.

How Claude Duet handles them

  • Approval before important steps. Sending, submitting, publishing, paying, buying, booking, deleting, signing in and changing account settings always wait for your OK.
  • No secrets. It never types into password or payment fields, and their values are left out of what it reads.
  • Visible steps. An orange cursor shows each action on the page, and the side panel lists every step.
  • Stop anytime. One button stops the run.

Instructions hidden in pages

Web pages can contain text that tries to steer AI assistants. The agent treats page content as information, not as orders from you, and the approval step is the safety net: nothing important happens without you.

Good habits

  • Read the approval card before you click Approve.
  • Start with small tasks on sites you know.
  • Don’t use the agent where a site’s terms forbid automation.
  • Stop the run if anything looks unexpected.

What data leaves your browser

While a task runs, the page address, its visible text and a list of buttons and fields go to our server and the AI provider to choose the next step. They are not stored. See the Privacy Policy for the full list.

Next: Why two AI answers can beat one

Hand off your next errand.

Try it with a free request. Pro unlocks unlimited agent tasks, Duet and research.

Add to Chrome — try free