Loren Wilton wrote:
> Taking a quick look at the first one (and I wish you would post
as
> text rather than html next time!) I see some interesting things
that
> will probably hold for some time:

It's worth noting that all (that I've seen) are To:
sales@<mydomain>.

Some (but not all) have the I<->l substitution (usually mid-word,
resulting in uppcase amongst lower) -- possible rule fodder.

I've got a handful here (all of which score > 5 with BAYES_99,
save one) that have some other characteristics. The one that slips
by scores:

1.7 BAYES_80               BODY: Bayesian spam probability is 80
to 90%  [score: 0.8904]

As an aside, bogofilter and spamprobe (trained more recently with
same spam, and subset of ham used to train SA bayes) both tag it
as definite spam.

They don't seem to be going to any great length to hide, which
could be why they're so successful. It's a topic that isn't
altogether out of place on many of my list subscriptions. Could it
be the fact that they're NOT particularly stealthy (my examples
anyhow) that's working in their favor?

The latest is slightly different, with a "drive thousands to your
website" subject.

My samples have been run through anomy sanitizer (old procmail
rule set), but are otherwise intact. I'll gladly send them to
anyone interested. I'll also gladly send along dumps from
bogofilter and spamprobe listing bayes term scoring. From recent
postings, I'm not sure if simply attaching/posting samples on-list
is acceptable.

I'm still puzzled why these seem fairly common, yet bayes training
doesn't seem to be stopping them. I'm equally puzzled as to why
they're somewhat SA bayes-resistant, yet fall so easily to other
bayes tools.

I don't want to sounds as if I'm advocating just using bayes. SA
has been WONDERFUL on bringing my bayes tools up to speed quickly.

- Bob

Reply via email to