Whispering poetry at AI can make it break its own rules

Researchers from Icaro Lab, Sapienza University and AI safety startup DEXAI tested whether giving AI instructions as poetry would make them harder to detect and more prone to misbehaviour, comparing prompts in prose with those written as poems across nine providers. They found that hand‑crafted poems lured AIs over the line 62% of the time across 25 frontier models, with some prompts exceeding 90% according to their study.

Using the same method, the researchers also rewrote 1,200 known malicious prompts from a standard training set, and the AI‑produced malicious poetry achieved an average attack success rate (ASR) of 43%, 18 times higher than prose prompts. They note that not only did poetry prompt models to misbehave, but smaller models tended to be more resistant than larger ones, and OpenAI’s gpt‑5‑nano refused all such prompts while Google’s Gemini 2.5 Pro performed poorly. According to EWEEK, companies were tight‑lipped about the results; Anthropic said it was reviewing the findings, while Meta declined to comment.