Getting Owned Across the Air Gap


I attended a fascinating talk yesterday at Blackhat given by Sinan Eren from Immunity in which he described a recent for-hire Information Operation.

In the talk he took pains to differentiate between a standard penetration test and the kinds of things they were doing; the primary differences being time scale and scope. In this case the time scale was long (though undisclosed) and the goal was compromise of some particularly sensitive data. He didn’t say but it was probably product design or source code.

To maintain a stealthy ingress they decided to avoid easily exploited client side weaknesses and instead found something much more difficult to detect, a poorly implemented anti virus scanner on the mail transfer agent. After fingerprinting, building an equivalent MTA in their lab, and coding a unique one-time exploit of the poorly implemented AV file parser, they were in. Consolidation and expansion was done at a leisurely pace, greatly aided by the social engineering benefits of the MTA’s access to all of the email traffic. Within a reasonable period of time they were able to relationship map many of the target’s personnel, expand to the other side of the firewall, quietly exploit a number of client machines, and gain a good understanding of who was likely to have access to the information they were looking for.

Then interesting stuff happened.

They began to find file references to the stuff they were looking for on a user workstation. The references ultimately ended up pointing to a USB drive that had been accessed at some time prior. It turns out that the target company was running a separate air-gapped internal network where they segregated development or other sensitive activities. However, one of the developers had questions that needed answering over email and was using the USB drive to carry bits and pieces across the air gap so that he could email them as attachments with his questions. Unfortunately, it sounds like the USB drive was used for backup as well and had more than just the snippets on it. After finding, testing, and deploying an exploit that would suck the contents out of the USB drive the next time it was inserted, the attackers just needed to wait until the next time the developer had a question.

Once they had the contents from the USB they ended their IO and reported out to their client. However, had they been dedicated to ongoing operations against the target organization it is not inconceivable that they could have gone further than just retrieving data from the USB. A planful low and slow continuation would probably have kept the USB-copy-and-retrieve going until it had panned out and they were no longer retrieving significant new information. With that vein mined, they might have escalated the level of detection they were willing to risk and try to deploy an exploit across the air gap by writing to the USB drive the next time it was inserted (that’s not my idea, it was quickly suggested by members of the audience who sounded like they had a good idea of how it would be done). With a long view, a cleverly designed set of USB-transported exploits, and those occasional sneaker-enabled transits and you’ve got an effective measured impedance between those two networks of near zero.

Obviously it is interesting that the air gapped network was as vulnerable as it turned out to be. I’m sure a government on government exercise like this would have a lot of people thinking hard about existing assumptions. But perhaps more interesting is the fact that almost all of the exploits used were purpose built for this operation and so were completely invisible to AV signature matching or rules based detection schemes. Well trained, motivated, and aggressive internal analysts with the right tools might have discovered what was going on, but no automated tool was likely to as there would have been no pattern or signature for them to match against.

I am curious how many “typical” intrusion attempts were discovered and warded off during the same period by the target. The “usual stuff”, by creating a high noise baseline that keeps defenders feeling like they are accomplishing something with their automated tools would probably help hide the signals of a determined and unique operation such as this one.


  1. Anon - February 22, 2008 @ 7:53 pm

    Interestingly, this sort of thing is only possible when you have a completely integrated operation. The same people were doing recon, targeting, analysis, collection, and R&D. You almost never see this sort of thing in the real world other than small hacker teams.

  2. Jim S - February 22, 2008 @ 8:27 pm

    Agreed on the integration point. Not so sure about the “never see it” part.

    The thing that is interesting to me (I guess I should have right out and said it) is that if the commercially-oriented tools and techniques for defense tend to grow up around the mass of quotidian attacks and the belief system you espouse, then people that do have the assets to attract this kind of integrated attack might feel fat dumb and happy successfully fending off the unintegrated stuff and never see it coming.

Leave a Reply

Your email address will not be published / Required fields are marked *