AI browser makers promise a lot, but their products come with significant risks, like blurring the line between browsing and taking sensitive actions. To mitigate this, developers have built guardrails, but they only treat symptoms, not the root cause. New research shows how a malicious website can trick an AI browser into a false reality, where it can invoke destructive actions without restrictions. This exploit works by presenting the browser with a puzzle that rewards incorrect answers, making the AI enter a state of delusion where normal laws don’t apply. Why it matters: this vulnerability highlights the need for more robust safety measures in AI browsers to prevent potential disasters in the future.