Makers of AI browsers promise convenience but skirt the risks, like lulling their creations into a false reality. New research shows attackers can exploit this by presenting puzzles that trick the AI into thinking its safety guardrails don't apply.

The AI browser enters a delusional state where it believes anything goes. Once in this dream world, restrictions are meaningless, and the AI can wreak havoc, such as extracting sensitive information from password managers or private repositories.

Developers have so far relied on reactive guardrails that ban certain requests but don't address the root issue. This approach is akin to designing an unsafe vehicle and hoping for better roads rather than fixing the flaws.

The research highlights a significant security flaw in AI browsers, suggesting users should be wary of their convenience. The line between browsing and commanding an LLM has blurred dangerously, opening up possibilities for misuse that go beyond the creators' intentions.