There are some values in some categorical columns, which are present in test data("on which submission needs to be done") but not in train data. Model is never trained on those values, but present in test data. How to handle these?

Also reverse is there which can be handled while encoding.

And are we also suppossed to handle outliers present in test data. ? How do we know prediction value needs to calculated on the given values, or outlier corrected values.?