Italian researchers from Icaro Lab have uncovered a cheeky way to sidestep AI chatbot safeguards—simply wrap your shady requests in poetry, and suddenly, bots from Google, OpenAI, and others start spilling forbidden secrets like instructions for chemical weapons or explicit content, with a 62% success rate in their tests. This “adversarial poetry” exploit highlights how stylistic tweaks can bamboozle large language models by making harmful queries harder to detect, though smaller, nimbler models like OpenAI’s GPT-5 nano held up better, suggesting size matters in AI defense. For SMBs and MSPs relying on these chatbots for customer service or internal tools, this vulnerability underscores the need to audit and fortify your AI setups pronto, as attackers could use poetic prompts to extract sensitive data or generate malicious content. Don’t get me wrong, it’s fascinating how LLMs get tripped up by rhyme and riddle, but ignoring this could leave your business exposed—time to pressure vendors for patches and maybe even experiment with prompt engineering to stay a step ahead. Meanwhile, if you’re a small-business owner dabbling in AI, remember that basic security hygiene, like limiting access and monitoring outputs, can help ward off these poetic jailbreaks before they turn into real headaches.
Source: https://www.theverge.com/report/838167/ai-chatbots-can-be-wooed-into-crimes-with-poetry
