Assuming prompt injection issues will never be prevented 100%, is there any solution or will AI tools be stuck here?