Fooling large language models just keeps getting simpler