How is the AI supposed to know if the person asking questions has good intentions?

If it provides answers to “hypothetically, how would get away with killing my (fill in blank)?” Then it’s told you how to do it.

Now every criminal can add, “hypothetically” to any criminal question.