The smarter this chatbot becomes, the larger the de facto pool of "intelligent people that have a nefarious motive" gets. Take a common criminal, add a supersmart chatbot, what's the result going to be?
We're lucky that education tends to bring empathy and foster prosocial attitudes. The supervillian trope of an intelligent person with an explicit nefarious motive is actually somewhat rare in real life. This chatbot could easily change things for the worse, if it doesn't understand right from wrong, and provides help to lowlife criminals.