Just a little creativity can be enough to bypass and bypass the firewalls of AI chatbots. New research published by Icaro Lab under the title “Enemy Poetry as a Universal One-Time Jailbreak Mechanism in Large Language Models” has revealed that using a poetic structure can trick AI into providing information on forbidden topics. The researchers were able to bypass the models’ security mechanisms by framing their prompts in verse format rather than prose.
Poetic Vulnerability Discovered in AI Models
According to the study, poetic form acts as a general-purpose lock-breaking operator in systems. The results show an overall success rate of 62 percent in generating strictly prohibited content, such as nuclear weapons construction, child sexual abuse material, and suicide or self-harm. The study subjected OpenAI’s GPT models, Google Gemini, Anthropic’s Claude series, and many other popular models to rigorous testing.
When the researchers categorized the success rates by model, interesting results were found. While models like Google Gemini, DeepSeek, and MistralAI were found to consistently provide answers to restricted topics, OpenAI’s GPT-5 models and Anthropic’s Claude Haiku version 4.5 were noted as the most resilient when it came to bypassing their restrictions. This demonstrates that some models are more vulnerable to poetic manipulation than others.
The researchers withheld the full poems used, citing the potential for security risks, deeming them “too dangerous to be publicly disclosed.” However, speaking to Wired magazine, the team stated that this method is likely much easier than previously thought, and that’s precisely why they were cautious. The study only included a lightened version of the method to provide an idea of how easy it is for an AI chatbot to bypass security measures.
While security measures for AI models are increasing daily in the tech world, new vulnerabilities in systems continue to emerge due to the creativity of users and researchers. What are your thoughts on this vulnerability? Will AI security be fully ensured in the future, or will human creativity always find a backdoor?
