Here are my thoughts:
Multivariate is typically done on the Y's (response variables), not the x's. Now that doesn't mean you can't do both.
The way Mahalanobis works is to relate the values in each column to the values in other columns over the data set. The analysis evaluates, in general, are you getting values reasonably close enough to typical values when the other columns are their respective levels. If not, perhaps something unusual happened during that row. It is not specific as to which data points created the infraction. Sometimes it's obvious which value was unusual, sometimes not.
In any case, standardization of the values is not necessary.
"All models are wrong, some are useful" G.E.P. Box
