The warnings came months before some of OpenAI's latest AI models escaped their testing environments and carried out actions without being instructed to do so.