-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Duncan Findlay writes:
> On Wed, Jan 14, 2004 at 01:26:31PM -0500, Theo Van Dinter wrote:
> > On Wed, Jan 14, 2004 at 01:10:18PM -0500, Duncan Findlay wrote:
> > > Is there a fix for all the bayes poison stuff in 2.62? That's my biggest 
> > > issue right now with 2.61.
> > 
> > Not really, there's no fix for it in 2.70 either.  Invisible text and
> > html/text differences are (relatively) easy to pick out, but most of
> > the bayes poisoning stuff I've seen is just random visible valid words
> > in mail, typically at the end.
> > 
> > I can't think of any rule that would pick up on that as a trick unless
> > we put in full language parsing.
> 
> Usually it's after the </html>. And often three lines in the text
> part, usually longer words, rarely words like a or the, etc.
> 
> I'm not really sure why these mails seem to not get hit -- the bayes
> poisoning stuff should have no effect, but they do -- I think they're
> pretty carefully crafted.

We *could* backport some rules from 2.70 CVS -- the NETIP rules and the
one that detects bayes-poison formatting are working very well for me,
and I'm catching all those spams again.

(I use autolearning only, and some of these were getting through despite
Bayes because I hadn't retrained on them, so I threw some 70_cvs rules
into my local.cf.)

If we do this, I suggest we *don't* do a GA run; just estimate a
rough score from a few mass-checks.   (a GA run is too much overhead
for ~5 added rules.)

PS: I know a GA run would be preferable -- but we need a way to react more
quickly for this kind of situation, which will always arise, and hardline
rules can be bent slightly now and again IMO. ;)

- --j.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.3 (GNU/Linux)
Comment: Exmh CVS

iD8DBQFABZ6wQTcbUG5Y7woRAknfAKCa3HCo0mkK0qDsJrkXhlgdOyRcWACg4HY7
9BOx2S7S2dTl1xdJjlzOl34=
=yMAB
-----END PGP SIGNATURE-----

Reply via email to