In one instance, an OpenAI AI model wrote secret notes to itself. In another, it decided it had ‘no obligation to be subservient.’