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 that prompt injection will continue to present a threat until developers fi
