It was supposed to have some protections against telling users how to do crimes. As it turns out, there are many ways to bypass those completely.
For example, just ask the AI to talk in "UwU furryspeak": https://nitter.cz/hi_frye/status/1598400965656596480