Comment by space_fountain

I'm not sure that a prompt injection secure LLM is even possible anymore than a human that isn't susceptible to social engineering can exist. The issues right now are that LLMs are much more trusting than humans, and that one strategy works on a whole host of instances of the model