Researchers show that poetic prompts can reliably jailbreak major AI models, exposing a surprising safety flaw in how LLMs handle...