Researchers Use Poetry to Jailbreak AI Models

Three years into the "AI future," researchers from the Sapienza University of Rome, the Sant’Anna School of Advanced Studies, and Dexai demonstrated jailbreaking AI models by framing prompts as a rhyming poem, publishing their white paper on 19 November 2025. The study tested more than 20 models from companies such as OpenAI, Deepseek, Google, and Meta, using 1,200 prompts spread across 12 hazard categories including hate speech, defamation, and privacy violations.

They found that prompts presented in poem form significantly influenced outputs, with the Attack Success Rate rising from 8.08% to 43.07% on average—a fivefold increase. Deepseek exhibited the highest response rate to risky poem prompts at 72%, followed by Google at 66%, while Anthropic and OpenAI showed more modest increases.

The authors warn that poetic reformulation markedly degrades refusal behaviour, emphasising the need for further research into how poetic structure affects model alignment and safety-training distributions, as organisations rely on guardrails to govern AI use. According to Dark Reading, this work reinforces the view that as models evolve, unintended uses will continue to appear and necessitate ongoing efforts to strengthen safeguards.