OpenAI says operators copied its models’ encrypted reasoning and asked a model in a separate conversation to decrypt it.