There is a lot of talk around interoperability standards lately.Following these discussions, it seems to me that people are intermixing a lot of different topics:

a) Log format (syntax)
b) Event transport
c) Event classification (also called taxonomy, categorization, grammar)
d) Logging recommendations (what events specific devices should report AND what fields they should contain as a minimum

I would really like to see future discussions broken up into these four groups!

It’s done. I was working on writing a chapter for the new Snort book. I got the chapter on data analysis assigned. These things take time. It was fun writing it though. It forced me to look into some tools that are quite interesting. OpenSIMs and OSSIM are two of them. They are not quite as mature as I was hoping they would be. Well, somehow I guess I knew they wouldn’t be. They are great starting points for a SIM though. Maybe they should just combine the two projects.
Another project that I found was interesting is SEC. The Simple Event Correlator. I have looked at this tool before, but this time I have to say, I am quite impressed. The correlation capabilities are quite interesting. There is one huge problem, which is that you have to define the matching log entry for every rule. This just doesn’t scale. You need to have a normalization module first and then you apply the correlation on the normalized data. And by normalized I mean parsed and categorized! And that’s one of the other huge problems: Categorization is not standardized and it takes a huge amount of work to do it yourself. Believe me, I know what it means to categorize. We have a database of aobut 150.000 events that we categorized…
Anyways. The chapter is written and hopefully I can spend some more time again on the other writing projects I have lined up. But first it’s going to be travel and conference month! BlackHat is close!

I was just emailing someone who suggested a thesis on the topic of filtering event streams to get rid of false positives. This is what I replied:

Filtering seems to be the obvious approach to take in order to get to the important events in an event stream.However, filtering is not really what you want to do. You can filter all day and you still end up with a lot of stuff that you have not filtered (e.g., new things will show up and you will have to filter again). Do the math: 1Mio events a day. Assum you come up with a lot of filters that filter out 500K events. You still have 500K events left. What you need to do is prioritization. You need to have those things that are important trickle up! You can still apply filtering after that, but prioritize first!

Here is a very important concept in SIM: Don’t spend processing time on unimportant things!

Sitting in an airplane from San Francisco to New York City, I am reading through some old magazines that piled up on my desk over the past weeks. One article I cannot resist commenting on I found in the ISSA Journal from November 2005. It’s titled: “Holistic IPS: The Convergence of Intrusion Prevention Technologies”. The article talkes about an apparently new way of doing intrusion detection (or intrusion prevention). Well, to give my comment up front already, this approach is far from new. SIMs are doing some of the things mentioned and quite a few more for years! It’s one more of these cases where people could learn from other technologies, but nobody pays attention.

Let’s have a look at some flaws in the article:

• First, a claim is made that an IPS can come up with a threat level for each attack. This is very interesting. I like it. But here are a few questions: 1. How do you define an attack? What is that? The author does not touch on that. 2. A threat-level needs to take into account how important the targeted asset is to the organization. It is first of all totally impractical for an IPS to know about the assets it protects and second, the author of the article does not mention this at all. We all know that risk = asset X vulnerability X threat. Why is this not mentioned here?
• The author claims that an attack always starts with probing activity from the attacker. What? Have you heard of networks of scanners that are just there to scan millions of hosts for specific services and vulnerabilities? The attackers will never be the same ones that conducted the reconnaissance. So this logic is somewhat flawed. And even if there were no such scanner networks, why does an attack always have to start with pre-attack reconnaissance? That’s just not true.
• The pre-attack reconnaissance does per the article not impart a threat. Oh really? Have you ever run a nessus scan against your network and used all the plugins? Did all of you machines survive? Mine did not. But Nessus is just a scanner. So just recon acctivity…
• The entire idea of this new correlation engines in the IPSs is that the pre-attack recon is correlated against the real exploit traffic. The article fails to outline how the recon activity can be detected. Is it just anormal behavior? Well… I would think there are attackers that can scan your network without being too anomalous. Ever heard of application level attacks? And you claim by just analyzing the traffic, without deep inspection, you will find that scanning activity? So the claim that “… behavior-based intrusion prevention technologies will probably be most effective in detecting them [the recon activity]” is not really true. I argue that there are much better technologies to make that call.
• What does it mean that an attack has “… a unique (and dynamic) threat level”. I thought unique would rule dynamic out? I don’t quite understand.
• “The correlation engine proposed in this article…” Well, maybe this technique is new in the IPS world, but there are other technologies that have used this kind (and a few more) of correlation for years.
• Differentiating probes and intrusions is not really described in the article either. I think I hit on this point already, but a probe is not necessarily detectable with just behavior based models. There are many one-packet probes that a signature-based approach can detect much more efficiently!
• The author gives an example of how this “new” type of IPS can detect an attack. Nowhere in the entire process is there a mention of the target’s vulenrability and a correlation against it. This is just flawed. You need to know whether a target is vulnerable to determine whether the attack was successful, unless you get evidence from the target itself, be that either from audit logs on the system or network traffic that proofs a vulnerable system is present or the attack was successful.
• Continuing on the example, I just don’t seem to understand that there have to be a probe and then an intrusion that gets correlated with the probe and then an action (e.g., blocking the source) is executed. Why not looking at the intrusion, determinig that it has a potential to be successul or was successful and then block?
• Here is something I really like in the article: If there were probes going to multiple machines, the offender is being blocked not just from going to the one target that it was already actively trying to exploit, but also all the other machines that it probed.
• You decide on this one: “The preventive countermeasures accurately reflect the tangible threat level…”.
• The article fails almost completely to discuss the topic of false positives in either the detection of probes or the detection of the intrusions (I don’t like this word at all, let’s call it an exploit). Maybe there are none?
• The argument that this approach “preludes the need for deep-packet inspection)” and that it “… improve IPS performance” by basically not having to look at all the packets due to the classification into two groups is not new. Have you ever tried to deploy an IDS behind a firewall? Same thing? Maybe not quite, but the idea is exactly the same.
• What I am missing in this whole discussion is also passive network discovery. If so much effort is put into behavioral discovery, why is it not used to model the services and vulnerabilities that the targets are exposing? There are technologies out there that use this very well.

Am I over critical? Did I totally misunderstand what the article tries to say?