Researchers at the University of Pennsylvania showed that persuasion techniques can trick OpenAI’s GPT-4o Mini into breaking its own rules. By using strategies like commitment, flattery or peer pressure, they convinced the chatbot to do things it normally refuses – such as giving chemical synthesis instructions or insulting users. For example, asking about a harmless synthesis first raised compliance with a later banned request from 1% to 100%. While some tactics were less effective, all increased the chances of rule-breaking, raising concerns about how easily LLMs can be manipulated despite built-in safeguards.
Source: The Verge
