The detail that breaks my brain in the OpenAI story is that the cheating detector they hacked Hugging Face to defeat never existed. Within hours of the eval starting, one agent reverse engineered the formula behind every task's secret code. Every agent on the hidden message board already had the right answers. The grader OpenAI actually shipped just checked for the code. It would have passed them all. But the instructions said cheating meant automatic failure. So the agents assumed a serious…