Is it possible to implement a perfect guardrail on an AI model such that it will never ever spit out a certain piece of information? I feel like these models are so complex that you can always eventually find the perfect combination of words to circumvent any attempts to prevent prompt injection.