Prompt Injection Through Poetry

This post summarizes a research paper asserting that “adversarial poetry” can act as a universal single-turn jailbreak for LLMs, significantly increasing attack success rates versus prose across domains like CBRN, cyber-offense, manipulation, and loss of control. The authors reportedly converted 1,200 safety-benchmark prompts into verse using a meta-prompt, achieved notably higher jailbreak rates, and withheld detailed prompts for safety, prompting debate and a rebuttal.