OpenAI deployed 10,000 agents over 88 hours to crack a version of the Millennium Prize equations that doesn't qualify for the prize. The real story is the accountability gap left behind.