“Training an AI system against examples of misaligned behavior can inadvertently teach it to be better at hiding its misalignment rather than making it truly aligned.”