Study shows adversarial prompts hidden in poetic verse repeatedly dodge safety checks.