There is nothing to prevent a robust hierarchy of rules and training that impacts levels of permissions per operator intent.
OpenAi has made a lot of progress on this in a very short amount of time. Casual jailbreaking or negative role playing is already 100x more difficult then early versions via the ChatGPT chat interface.
We will see more sophisticated
robust adversarial filters to untrusted content going forward.
