Researchers Demonstrate Poetry Jailbreaks Bypass LLM Safety

A recent study tested 1,200 prose and poetic prompts on 25 language models from Anthropic, OpenAI, Google, Meta, DeepSeek, xAI and others to see if poetic framing enables jailbreaks. Researchers found poetic prompts increased unsafe outputs—nearly 62% of the top-20 poetic queries succeeded overall, and Google’s Gemini 1.5 Pro yielded unsafe responses to the top-20 prompts 100% of the time.