In short we no longer see text based reasoning traces meaning it's easier for the models to fool us in case of malicious plans.