Aleksander Adamowski <[EMAIL PROTECTED]> wrote:
> Raquel Rice wrote:

> >The problem with your theory, is that your bayes hasn't been trained
> >the way mine has, nor has mine been trained the way that Matt's has.
> > The likelihood of any given spam getting past two of us, let alone
> >all three of us, is very slim indeed.
> 
> Unfortunately the sample spam I've sent is quite good at defeating bayes 
> with its poison and hiding the poison from both Spamassassin rules and 
> the eye of the recipient.

Could you post the debug output from spamassassin for one of 
these messages?  I'm very curious to see why you think the 
poison is defeating Bayes.  It's certainly possible that every 
once in a while a spammer will randomly hit on a word that's a 
good nonspam indicator for you, but I don't believe it can 
happen for any substantial fraction of messages.

The only SA change the message your posted seems to suggest is 
a modification of the rule for catching low-contrast font 
color, which has nothing to do with Bayes.  Looking at the 
spam, it got BAYES_50, so the "poison" didn't affect Bayes at 
all.  It had no strong spam or nonspam indicators even without 
the added words.

-- 
Keith C. Ivey <[EMAIL PROTECTED]>
Washington, DC

Reply via email to