OpenAI has disclosed six reports of "unexpected or concerning" behavior in artificial intelligence models as the debate on AI safety becomes increasingly heated.