In a discovery that could reshape how the tech world thinks about AI security, a new study by Anthropic has revealed a surprisingly simple method for compromising large language models (LLMs).