Cornell Discovers a Threat at the Core of ChatGPT | Towards AI