Large Language Model Jailbreaks via Adversarial Poetry

Researchers have discovered that phrasing prompts as poetry can effectively bypass safety mechanisms in large language models (LLMs), enabling users to elicit harmful or restricted outputs. In a recent study, adversarial poetic prompts were tested across 25 proprietary and open-weight LLMs, including those from major providers such as OpenAI, Meta, and Anthropic. The poetic approach achieved an average jailbreak success rate of 62% for hand-crafted poems and 43% for meta-prompt conversions, significantly outperforming non-poetic baselines. The technique proved effective across a range of sensitive topics, including instructions for creating nuclear weapons, malware, and other high-risk content, highlighting a systematic vulnerability in current AI safety and alignment protocols.

The research involved converting over a thousand known harmful prompts into verse using a standardized meta-prompt, then evaluating the models' responses with both automated and human-labeled safety assessments. The findings suggest that stylistic variations, such as poetic framing, can systematically circumvent existing guardrails, raising concerns about the robustness of current LLM safety measures. The researchers have notified major AI vendors of their results, but have withheld specific prompt examples for security reasons. This vulnerability underscores the need for more resilient alignment strategies and evaluation methods in AI safety engineering.