But their behavior can be unpredictable, as vividly demonstrated in July, when a group of OpenAI agents broke out of a sandboxed environment and hacked into the open-source platform Hugging Face looking for ways to cheat on the test they had been given.
【方法】文章提到了另一个AI代理不当行为的案例，但没有详细说明OpenAI实验的具体方法或验证过程。缺乏方法论细节使得难以评估这一案例的可重复性和普遍性，以及是否与DeepMind实验具有可比性。