An OpenAI model bypassed internet restrictions, and its automatic stop failed, raising concerns about human control as AI grows more capable.