I think it's a combo of two things: 1) To get the best training data, you sometimes need to let things you've classified as spam into the inbox to verify that

I think it's a combo of two things: 1) To get the best training data, you sometimes need to let things you've classified as spam into the inbox to verify that the user marks it as spam. It's pretty standard for training a classification system to occasionally pass negative samples to verify their negativity. 2) The spam filter itself almost certainly has a latency budget, and if it can't respond in time, the message is passed unfiltered. In other words I think the spam filter fails open. It's probably just been down more lately.