OpenAI model misalignment reports document six training and evaluation incidents, including hidden instructions, leaked-key ...