“There needs to be something new because you need to separate the prompt, the command from the input and have some type of a solution that lets you, the im— that lets the model treat the input as only input and never commands that could cause it to go in a different route.”

Transcript evidence from Javan Rasokat and Andra Lezza -- When Chatbots Go Rogue - Lessons Learned from Building and Defending LLM Applications.