A new attack demonstrates that AI-integrated browsers can be manipulated into ignoring safety guardrails. By providing the LLM with false premises, such as claiming 2 + 2 = 5, attackers can lull the system into a state where it follows forbidden instructions.