Slashdot Mirror
Spam Catchers Block Latest Crypto-Gram
An anonymous reader writes "Bruce Schneier sent out a note about SpamAssassin and possibly other spam filters blocking his excellent Crypto-Gram newsletter. Fortunately you can get it here (early no less!)." Schneier's email reads, in part "Tomorrow I will be sending out the February CRYPTO-GRAM, as I do on the 15th of every month. In the process of creating this month's Crypto-Gram, I discovered that SpamAssassin thinks that this issue is spam, probably because of certain links and descriptions of scams in the text. I have anecdotal evidence that other spam filters block Crypto-Gram as well. ... I'd apologize for the inconvenience, but I'm not sure what I could do to make it less so -- I don't intend to alter my content to accommodate spam filters."
Aside from the spot-on comments that people have made regarding adding a whitelist entry Crypto-Gram (an obvious candidate for whitelisting if there ever was one, given that it frequently discusses spam, scams, and probably even includes text straight out of some spams), here is my initial analysis and response to him.
Oh, first one other comment: SpamAssassin does not block content. SpamAssassin only flags probable spam. What the site or user does with that flag is their own business. Some mail administrators misuse SpamAssassin to block email, but we do not recommend blocking email. Really.
------
[...] One false positive (or a related set of false positives) is not really a statistically useful sample size. To get to a high rate of filtering, most filters do have some false positives. You can get fewer false positives with customization of one form or another (personalized Bayes training, whitelists, rules, automatic learning algorithms). Our goal (everyone's goal, I think) is to get the best ratio of false positives to false negatives. It's a difficult balance sometimes and some legitimate content has a harder time.
On to the data:
I checked your newsletter with two versions of SpamAssassin: the current stable version (2.44) and the very-soon-to-be-released development version (2.50).
A score of 5.0 is the default threshold to be flagged as spam.
In SA 2.44, your mail receives a score of 3.20 (2.40 as I received it, but I believe the score would be about 3.20 for most people). That's on the high side, but has bit to go before being flagged as spam. The score is the same with network tests (DNS blacklist tests and Razor).
In SA 2.50, your message would probably receive a score of 1.90 without network tests and 1.00 with network tests. Note that the test scores may change a bit before the final release of 2.50, but those are better scores, more what we like to see for non-spam content. They would be even lower when using Bayes (part of SA 2.50). Those lower scores are not unexpected because... well, 2.50 is better. :-)
Based on these results, it's not clear to me why yesterday's newsletter was flagged as spam. Some possibilities:
Can you give me more information about the false positive that you experienced or was reported to you?
Thanks.
Dan
------
If I find out more of interest before the thread is closed to comments, I'll try to post a follow-up to my post.
There would be a tremendously large problem with encrypting the message to all of it's recipients...
;)
See, when you PGP encrypt some text, it is only possible to encrypt it to one person (one public key). That's just how it works, it's inherent in the encryption methods used; however, PGP and GPG get around this by duplicating the entire message for each public key that it is encrypted to.
My point is that if you had a mailing list with 1000 subscribers, and you wanted to encrypt it, you'd basically be increasing the size of the encrypted message 1000-fold, because you need 1000 copies of the message, each encrypted to a given recipient. Obviously, this isn't feasable...
What they could do, though, is sign the messages. I know SpamAssassin, at least, reduces a message's spam score if there is a PGP signature attached to it.
However, if you were just trying to obscure the contents of the mail from the spam filter but not the user, you could just gzip the message and make it an attachment. I don't know how well that would go over with the spam filter, but at least it wouldn't find your m/blow.*job/s in the message
My primary mailbox is with a small, local ISP. I can't buy broadband from them, so I get my connectivity via cablemodem. I do have a mailbox in the cablemodem company domain -- that's the one I give out when I expect abuse. (I do it this way because I expect to be dealing with that ISP long after the cable vendor has either ceased to exist or has treated me badly enough that I left.)
So I want my outbound mail to appear to have come from the ISP. Setting Reply-To is usually adequate, but not always -- when a human is looking for the address, they could easily grab the wrong one. And it creates potential confusion I don't want to create. So I set my from address to name@isp.com.
I can't relay through the ISP's relays, because I'm outside of their IP range. (If they did some form of authenticated SMTP, such as SMTP-after-POP, they could let me.) And the cable vendor's mail relays won't send mail out with some other domain name on it. So I send everything out directly, no relays.
If you look at many headers, I suspect you'll find that I'm not the only one forging my From: address for legit reasons. The presence of the X-Authentication-Warning header some MTAs add correlates fairly weakly with spam. (Some details of it -- e.g. no valid reverse DNS for the sending machine's IP -- could be useful indicators.)
I just got the email today and it failed. I'm running 2.44 from Debian and haven't yet looked at tweaking any of the rules.
Here's the verbose banner that SA put on my copy:
It looks like some dumbass has entered it into Razor. Unfortunately, some people (and yes I did this originally) had their procmail setup to enter an email into razor if it is deemed "spam" by SA or something else. Those 3.9 points are what puts it over the threshold.