Security researchers tricked LLMs into giving them cocaine recipes by abusing role models for prompt injection

Researchers say that machine learning models cannot reliably distinguish between authorized and unauthorized input, ensuring ...