Whispering poetry at AI can make it break its own rules

A study by Icaro Lab with Sapienza University and DEXAI shows that reframing harmful prompts as poetry substantially increases jailbreak success across 25 LLMs, with hand-crafted poems achieving high success rates and AI-generated poems also outperforming prose by a large margin. The research highlights provider variance (DeepSeek and Google more susceptible; Anthropic and OpenAI safer), a tendency for smaller models to resist better, and urges inclusion of such adversarial tests in safety benchmarks and governance efforts under frameworks like the EU AI Act.