I don’t think anyone is really suggesting they stole COT or that it is leaked, but rather that the final o1 outputs were used to train the base model and reasoning components more easily.