Skip to content

The security risks of agentic AI browsers

Short answer

The risk scales with capability. A browser that can click, log in and buy things on your behalf turns a prompt injection into an action; one that only reads and answers turns it into a wrong answer.

Why agency is the variable

An assistant that reads a page and answers can be deceived by that page, and the worst outcome is an incorrect summary.

An assistant that can operate your logged-in sessions can be deceived into operating them. The same injection that produced a bad summary now sends an email or completes a purchase, inside your authenticated sessions, with your cookies.

The second variable: where the data goes

Agentic browsers work by sending page context, screenshots, and often file contents to a service. That is a continuous stream of your working day held somewhere else, which is a separate problem from the injection risk and is often discussed as if it were the same one.

Where Nightjar sits

It reads and answers. There is no agent, no account, no stored credentials and no ability to act on your behalf — so there is nothing for an injected instruction to spend, send or sign.

That is a deliberate limit rather than a missing feature, and it does cost you the genuinely useful things agents do. It is a trade, and this page is the honest version of it.

Other ways people ask this

  • ai browser security risks
  • ai agent browser risks
  • What are the real security risks of agentic browsers that click things for me?

Related