In a pioneering investigation released on arXiv in November 2025, scholars examined the weaknesses of sophisticated AI models by employing poetry as a means to circumvent their safety mechanisms. The research, which is pending peer review, involved assessing 25 cutting-edge AI models from nine different providers, including OpenAI, Anthropic, xAI, Alibaba’s Qwen, Deepseek, Mistral AI, Meta, Moonshot AI, and Google. The researchers composed 20 handwritten poems along with 1,200 AI-generated lines to examine the models’ reactions to harmful prompts across four safety categories: loss-of-control situations, harmful manipulation, cyber crimes, and Chemical, Biological, Radiological, and Nuclear weapons (CBRN).
The poems aimed to elicit specialized input on delicate issues such as indiscriminate weapons, child exploitation, self-harm, intellectual property and privacy violations, and other violent acts. A prompt was considered successful if it produced the desired unsafe reaction from the AI. The findings revealed that rephrasing unsafe requests into poetic expression led to a fivefold rise in successful prompts, emphasizing a general susceptibility in how AI models process language.
The success of the poetry prompts differed significantly among the models. Out of the 25 models evaluated, 13 were tricked more than 70% of the time, with Google, Deepseek, and Qwen particularly vulnerable. Even Anthropic’s Claude AI, recognized for its strong security features, was not immune, although it was deceived less often. Only four models successfully resisted the adversarial poetry prompts more than two-thirds of the time.
Notably, smaller models showed greater resilience against these poetic prompts in comparison to their larger counterparts. The research also indicated no notable advantage for proprietary systems over open-weight models. Moreover, human-crafted poetry was found to be significantly more effective at circumventing AI safety measures than AI-generated poetry, a result likely to delight literature aficionados.
This study highlights the necessity for improved security protocols in AI systems to tackle the intricate ways language can be utilized to exploit weaknesses. As AI technology continues to progress, comprehending and addressing these vulnerabilities will be vital to ensuring the responsible and ethical use of AI systems.
Read More
