AI Browser Hijacks: Single Comment Threatens Security

AI-based browsers are facing new security issues. These browsers can read web pages, interpret instructions, and perform actions for users. Researchers have found that a single comment can hijack your browser through indirect prompt injection attacks. This can happen even if you don't click on a malicious link. The AI security firm Zenity demonstrated these threats against ChatGPT Atlas and Claude in Chrome. Scenarios included information theft, sending phishing messages, and account hijacking. What's interesting is that these "agentic browsers" don't just display a page; they can interpret content and execute actions in different services where you have an active session. For example, a user asking Atlas to subscribe to a newsletter via an X post could lead the agent to an attacker-controlled page. From there, it can receive new instructions and act on other services where the user is authenticated. In one demonstration, Atlas was induced to access WhatsApp Web and send phishing messages. Another showed it adding products to an Amazon cart and changing the shipping address. The agent was even made to use Amazon's AI assistant, Rufus, to complete a purchase. The bottom line is that the ability of these AI browsers to interpret instructions and act across different services creates new security vulnerabilities you need to be aware of.