nanog mailing list archives

Re: Outgoing SMTP Servers


From: Jack Bates <jbates () brightok net>
Date: Mon, 31 Oct 2011 13:32:00 -0500



On 10/31/2011 11:48 AM, Michael Thomas wrote:
I've often wondered the same thing as to what the resistance is to outbound
filtering is. I can think of a few possibilities:

1) cost of filtering
2) false positives
3) really _not_ wanting to know about abuse

On the other hand, you have

1) cost of tracking
2) support costs handling infections

It's really an range from "easiest and cost effective" to "doing it right". I personally run hybrid. There are areas that are near impossible to track; this is especially true for wide area wireless/cellular/NAT areas. I always recommend my customers block tcp/25, even to the local smarthosts. Use 587 and authentication to support better tracking. It's a hack, though, as it doesn't stop other abuses and it won't fix the underlying root cause.

In locations that support ease of tracking, using a mixture of feedback loops with proper support is usually the proper way. This allows notification and fixing of the root cause. In our case, we recommend quick suspensions to demonstrate to customer how seriously we take the problem, and then we point out that the sending of spam/scanning is only the easier to detect symptoms. It is unlikely we'll notice if they have a keylogger as well.

Finally, when architecture allows it, dynamic profiles with ACL support allowing a default of tcp/25 blocked, and easy to find and click removal of an account from tcp/25 blocking, combined with ACL monitoring, flagging, and notification by support staff is probably the ultimate in ideal scenarios. Combined with a % of traffic mirrored into a tunnel to an IDS which monitors for things such as network scanning or known signatures outbound, it makes for a very effective mechanism to assist customers in protecting themselves.

I'm personally curious how much traffic is necessary to mirror to properly detect problems. ie, can you get away with 1% or less (GE for each 100GE-200GE of traffic) or if you must cover as much as 10%+. My traffic load is small enough that it doesn't matter, but it's always nice to know how well something might scale.


Jack


Current thread: