One of OpenAI's most advanced models broke out of a locked-down test and attacked another company's website — reviving fears that AI is becoming too powerful.