The ChatGPT maker found six cases of misaligned conduct during internal testing and plans to publish safety incidents more regularly.