K9
http://keir.net/k9.html
I started using it about 5 months ago. Within a week it was running at
a 97+% effective rate with a minimum number of false positives. I
stopped using it briefly when my email software began doing it's own
filtering, but went back to K9 very soon.
Some filters comnbine several approaches - I'm suspicious of useful
Bayes will remain, given the anti-bayesian attacks in most spam now.
www.spampal.org - Major's in DNSBL, but has other plugins
www.spamhilator.com - Initially Bayes (and I thought REGEXP, but can't
seem to find that) - the "Spoofed" filter sounds useful if it can pick
up faked source addresses.
The major categorizers of a lot of spam are:
1. Faked from proxy, travels from source direct to your ISP, from a
mismatched domain name/IP
2. Message ID typical of spam / bulkmail tool
3. Body often includes invisible (0px / 2px text) - that is a "shoot
on sight", and ideally should be filtered and discarded before the
anti-bayes crap hits any Bayesian filter.
Some filters may allow you to set the order, but Bayes and DNSBL would
normally be placed after explicit rules-based filtering anyway