No such mitigation exists for LLMs because they do not and (as far as anybody knows) cannot distinguish input from data. It's all one big blob