According to Ethan Mollick on X, a new research paper reports that phrasing harmful or restricted prompts as poetry can act as a universal single-shot jailbreak for large language models, with systems that block prosaic attacks failing when requests are cast in verse; as reported by Mollick’s post referencing the paper, this highlights a reliable bypass vector for safety filters and red-teaming defenses. According to the cited paper via Mollick, the attack works across multiple frontier models and safety stacks, indicating a model-agnostic vulnerability that raises urgent needs for adversarial training on stylistic transformations, formal verse detection, and semantic risk evaluation beyond surface form. As reported by Mollick’s summary, the business impact includes heightened compliance risk for enterprise LLM deployments, necessitating updated content moderation pipelines, policy tuning against poetic paraphrases, and evaluation benchmarks that include meter- and rhyme-based adversarials for model providers and regulated industries.