Mike Jackson wrote: > I've been getting quite a few spams (which slipped past SA) in the last > few minutes with subject lines like "dies in McDonalds", so I looked at > the message source to see how they were scoring (which I've included > below). In all the cases, the HTML content (at least as displayed in > Outlook Express) was fairly consistent, but the plain text version > looked like typical Bayes poisoning text. >
Really, I'd be looking into why the messages got past SA. Did it get a decent BAYES_ score? The bayes "poison" really shouldn't be a problem. The use of chi-squared combining makes bayes poisoning pretty ineffective as long as you're training your bayes often and training well. And by "training well" I specifically mean you must train spam messages containing "poison" as spam. If you're avoiding training "poison", then you yourself are making that poison effective. (Bayes can only be as accurate as its training. If its not getting realistic training, it won't do well with realistic mail.)
