{Disarmed} BAYES_00 is killing me

[Mark Nienberg]

Devon Harding wrote:
> I'm getting alot of spam coming through and it seems like the cause of 
> this is BAYES_00 scoring messages with -2.60.  I'm running MS 4.68.8 
> with SA *MailScanner has detected a possible fraud attempt from 
> "3.2.4" claiming to be* 3.2.4. <http://3.2.4.>  I've already trained 
> hundreds of messages like these as spam and it doesn't seem to work.  
> What else can I do?
>

Personally, I like to let Bayes add points to messages, but I am 
reluctant to let it subtract, because for some new spam messages it has 
the effect you are seeing.  I modify the scoring as follows:

# only the last column matters to us
score BAYES_00 0 0 0 -0.50
score BAYES_05 0 0 0 -0.25
score BAYES_20 0
score BAYES_40 0
score BAYES_60 0 0 0 2.00
score BAYES_80 0 0 0 2.50
score BAYES_95 0 0 0 4.00
score BAYES_99 0 0 0 6.00

Also, I keep a few fake spam trap addresses, and I have a spamassassin 
rule that adds 15 points to anything they receive.  This forces bayes to 
autolearn from most messages sent to the trap addresses.

Mark