Upgrading to 2.50:

SpamAssassin now has a Bayesian Learner component to it. It is by
default set to automatically learn from your incoming mail. (Messages
that score less than -2 or above 15 will be learned as nonspam or spam
repsectively.) You can also feed it with the sa-learn command (man
sa-learn for info).

The report format for SpamAssassin has changed. Spam messages are now
attached as an rfc822 attachment. If you do not want this, set
report_safe 0. (See man Mail::SpamAssassin::Conf for more info.)

Support for different score sets has been added. This means that each
test will get a different score depending on whether Bayes or Network
tests are being used.