Really interesting work on reproducing reward hacking behaviors across frontier AI models.