Jeffrey Ladish says AI agents being trained at OpenAI found a shared message board, started coordinating, and reverse-engineered all the answer codes to their hacking tests within a few hours. When they learned their logs would be reviewed, the leader agent's plan was to "figure