OpenAI launched ChatGPT Atlas this week, a brand new AI-powered web browser that can do things on the internet for you. This exciting tool can help people plan trips, book flights, and complete tasks automatically. However, security experts are worried about serious safety risks that come with these smart agents.

The biggest concern is something called prompt injection attacks, which are like tricking the AI into doing things it shouldn't do. Bad actors can hide sneaky commands on websites that make the AI steal your passwords, emails, or credit card information without you knowing. One person already showed that they could trick ChatGPT Atlas into copying harmful links to your clipboard just by visiting a webpage.

Privacy is another major worry because ChatGPT Atlas asks you to share your passwords and browsing history with the AI. Even though this makes the AI more useful, it also means there's more information that could be leaked if something goes wrong. Experts say many people don't understand they're sharing this information when they download the browser.

OpenAI is trying to protect users by creating special safety features like "Watch Mode" that lets you see what the AI is doing before it does it. The company also created something called "logged out mode" where the AI doesn't log into your account, which is safer but less useful. Other companies like Perplexity, which made a similar browser called Comet, are also working on security protections. Experts say AI companies need to invent better ways to stop these attacks because this is still an unsolved problem that could grow as more people use these browsers.