Researchers tested seven popular AI browsers and found four vulnerable to attacks that trick the AI agent into handing over personal data.