If you foresee this situation occuring, you could create a class 'OTHER' and populate it with something that would make sense, that could mimic well what you expect your model to exprience at test time. If your training set contains a lot of classes, good candidates for combining into the 'OTHER' class might be members of the least represented categories.