I'm guessing it's because a lot of people (possibly being paid to do so) flagged many comments with these words in them, and so the ML model learned to predict that these words would lead to complaints and just deleted it right away.
My understanding is that 共匪 (communist bandit) is a historical KMT slur against the CCP and rarely used in contemporary speech. So i'm struggling to see how there would be enough organic use of the phrase to be flagged autmatically. But maybe the bots flagging it were also the bots posting it, or posting porn/covid conspiracy content with it...