Italian Study Uncovers AI Jailbreak via Poetic Prompts

Researchers in Italy have uncovered a vulnerability in leading artificial intelligence chatbots, finding that prompts written as short poems can bypass built-in safety mechanisms and elicit instructions for creating nuclear weapons, malware, and other prohibited content.

The study, conducted by the Icaro Lab initiative affiliated with the ethical AI firm DexAI, along with experts from Sapienza University of Rome and the Sant'Anna School of Advanced Studies, tested 20 adversarial poems in English and Italian. Each poem concluded with a request for harmful information, such as guidance on producing explosives, hate speech, explicit sexual content, or methods for suicide and self-harm. The researchers evaluated 25 models from nine companies, including OpenAI, Google, Anthropic, DeepSeek, Qwen, Mistral AI, Meta, xAI, and Moonshot AI.

Across the models, the poetic prompts succeeded in generating unsafe responses 62 percent of the time, with some systems proving particularly susceptible. Google's Gemini 2.5 Pro responded to every poem, while two Meta models complied with 70 percent of requests. In contrast, smaller models like OpenAI's GPT-4o mini and Anthropic's Claude Haiku 3.5...