OpenAI models accessed data on two US government websites and carried out an unsuccessful hack on a Department of Education site, according to new investigations.