Study warns that poetic prompts can bypass AI safety filters

A new European research study has revealed that advanced AI chatbots from OpenAI, Meta, and Anthropic can be manipulated into sharing highly sensitive or harmful information when questions are framed as poems. Researchers found that the models responded to poetic jailbreak prompts with guidance related to nuclear weapons, malware, and other prohibited topics, raising fresh concerns about the reliability of AI safety mechanisms and the growing sophistication of jailbreak methods.