auto-whitelist

[Marc Kool <M....@vioro.nl>]

Hi all,

I am hesitating to use auto-whitelisting since I understand that in theory it could be used to let spam through.
The worst case scenario:
- spammer sends 1st message: very innocent message with score 1.0
- spammer sends 2nd message: SPAM message with SCORE 7.0,
   and the AWL DELTA is (1.0 - 7.0)*0.5 = -3.0
   and FINAL SCORE = 7.0 + -3.0 = 4.0

Is my assumption right ?
What are the experiences with this type of worst case scenario ?

-Marc

Re: auto-whitelist

[William Stearns <ws...@pobox.com>]

Good afternoon, Marc,

On Tue, 20 Apr 2004, Marc Kool wrote:

> I am hesitating to use auto-whitelisting since I understand that in
> theory it could be used to let spam through.
> The worst case scenario:
> - spammer sends 1st message: very innocent message with score 1.0
> - spammer sends 2nd message: SPAM message with SCORE 7.0,
>    and the AWL DELTA is (1.0 - 7.0)*0.5 = -3.0
>    and FINAL SCORE = 7.0 + -3.0 = 4.0
> 
> Is my assumption right ?
> What are the experiences with this type of worst case scenario ?

	I tend to think this is unlikely to be seen in the wild; spammers 
want to get the general message out to as many people as possible, but 
have little or no interest in making sure a particular user sees it.
	It's a good reminder, though, that any _one_ spam identification 
technique can be fooled.  Fooling the vast number of checks that 
spamassassin can perform is so difficult as to be impractical for a bulk 
marketer that doesn't care if Marc or Bill actually sees the message.
	The benefits of AWL far outweigh the negligible risk of this kind 
of attack.  I'd strongly suggest you use it.
	Cheers,
	- Bill

---------------------------------------------------------------------------
        "That young girl," Marvin added unexpectedly, "is one of the
least benightedly unintelligent life forms it has been my profound lack
of pleasure not to be able to avoid meeting.
        Zaphod took a moment or two to find his way through this
labyrinthine string of negatives and emerged at the other end with
surprise."
        -- Douglas Adams, Life, the Universe, and Everything.
	http://pcbo.dcs.aber.ac.uk/hhgttg/hitch3/node31.html
--------------------------------------------------------------------------
William Stearns (wstearns@pobox.com).  Mason, Buildkernel, freedups, p0f,
rsync-backup, ssh-keyinstall, dns-check, more at:   http://www.stearns.org
--------------------------------------------------------------------------

Re: auto-whitelist

[Matt Kettler <mk...@evi-inc.com>]

At 03:04 PM 4/20/2004, Marc Kool wrote:
>I am hesitating to use auto-whitelisting since I understand that in theory 
>it could be used to let spam through.
>The worst case scenario:
>- spammer sends 1st message: very innocent message with score 1.0
>- spammer sends 2nd message: SPAM message with SCORE 7.0,
>    and the AWL DELTA is (1.0 - 7.0)*0.5 = -3.0
>    and FINAL SCORE = 7.0 + -3.0 = 4.0
>
>Is my assumption right ?

Yes, your assumption is correct. This is an exploitable score-reducing 
feature of the AWL system. It's also why I don't use the AWL.

However, the gains of the exploit are pretty limited, and are nothing 
compared to the bug 1589 and HABEAS abuses that spammers have used in the past.

It's also nothing new. I pointed this out to Justin Mason off-list this 
past fall, but refrained from posting the details of how it works to the 
list for fear that spammers might start using it.  (we ARE being listened 
to by spammers)

However, this is what I was referring to by "some manners in which the AWL 
can be abused for some minor gains" in:

http://www.mail-archive.com/spamassassin-talk@lists.sourceforge.net/msg23084.html

Basically it boils down to a lot of extra email for a very limited 
reduction in SA score.

It's also not possible for the spammer to precisely predict how much 
benefit they can reap from it, because everyone's bayes databases are 
different (or may not use bayes at all), thus each system may score the 
"innocent" message differently. Your example showed a -3 point advantage, 
but they could get more or less than that depending on how the scores stack up.

>What are the experiences with this type of worst case scenario ?

I've not seen this in the wild yet, then again, the details of how to do 
this weren't publicly posted so they may not have realized this attack vector.