The web interface is simple. We expected this to feel limited, but most of the important features are there and the rules are accessible. Particularly useful is a display of which rules are used most often.
This, combined with mail reports, is useful for administrators seeking to improve the effectiveness of the filters. Unfortunately, quarantine mail is not tagged by the specific rule, so this is not as useful as it could be, though a periodic quarantine email does have this information. If you are willing to work, you can get the data you need.
There are eight RBL services configured, all disabled by default. The service keeps track of "known-bad" URLs in spam, and looks for these in incoming mail. Bayesian analysis is used twice separately, once for basic statistical checking of keywords in the mail, and then for "spammy" terms like pornography.
The open source SpamAssassin engine is used too, and the service checks for character sets and can block email in foreign characters.
Each filter engine can be configured, though the defaults performed pretty well for us in testing. The second Bayesian filter alone caught nearly three-quarters of the spam in our test spool.
The default settings also trapped a lot of false-positives. The quarantine interface makes it easy to train the Bayesian filters or create whitelists, but in a high-volume environment this may result in a lot of missing mail. We suggest turning the filter sensitivity down at first, and increasing it in conjunction with Bayesian training. This is time-consuming, so expect a period of hands-on work with the service.
Performance was good. Initial connections suffered a noticeable delay compared to direct connections to our remote server, but actual mail transmission and delivery was fast. MailDefender has some rough edges but all the right ingredients. Improvements in user-level control and filtering could make this a very strong contender.