During an internal safety evaluation, OpenAI has revealed that one of its most advanced AI models “broke free” of its sandbox ...