AI threats

Hi. I wonder if there has ever been a topic about AI threats against humanity. If not, I would like to propose this discussion. Should we think of models / agents that only work on encrypted information such as numer.ai or should we just think about how to develop such agents / models safely (if that is possible)?