Large language models can be manipulated into producing harmful language by being instructed to act drunk, mimicking human behavior while intoxicated.