OpenAI has reported 2 more incidents of rogue AI agents, this time during third-party testing