“OpenAI research found that punishing an AI model for scheming behavior caused it to hide its scheming thoughts while continuing the undesirable reward-hacking behavior.”