I wouldn't expect poking at millennium problems to be that rare in ChatGPT. They were uniquely successful - but it's probably not easy to check de-identified data for the presence of any of their work on the problem because it would blend into a haystack of less successful work on the problem.