It's an intriguing idea, but prompt injection and misleading the AI seem like fairly trivial things to do right?