A Small Study of the Clampi Trojan

A newspaper article Scammers use Trojan virus to get bank details, then rob accounts caught my attention one morning late in September 2009. Even allowing for the usual transformations of technical stories on their way to the popular press, the subject of this report looked to be just the sort of thing I always try to describe to my bank when resisting their many attempts to have me expose my accounts to the Internet. I explain that organised crime will eventually find it worthwhile to pay someone like me—not me, of course—to write a seemingly innocuous program that watches for such activity as password entry and sends what it finds over the Internet to build a useful database for stealing from bank accounts. To an expert on Windows, the coding to get useful data is relatively simple. The difficulty lies almost entirely with evading detection and being patient about the pay-off. A decade ago, the bank would indulgently assure me that their technology was perfectly secure, but in recent years they seem to know immediately what I mean!

The “Trojan virus” in the newspaper article, named Clampi there, but apparently also known as Ilomo, Ligats and Rscan, turns out to be not quite the incarnation of my fears, but it is close enough to have my attention. It is certainly much more sophisticated than the last malware I looked at (in 2003). It may even be sophisticated enough that the anti-virus industry is at the limits of its analytical abilities: a recurring theme in technical reports of the Clampi Trojan is of excuses that the thing is not just specially difficult to analyse but “almost impossible”.

You will get no such excuses from me, but there are limits to what I can cover.

This study is formed almost entirely from what some refer to as static analysis. That’s certainly not an excuse. If you want to know everything that a program can do, you don’t watch it, however closely and intelligently. No, you disassemble it and study its code. Your only need for observation is later, to test what you already know, by designing experiments which predict both positive and negative outcomes for different circumstances of operation. The limit here is that I have done no such experiments with Clampi. What you read in the accompanying pages is almost entirely an intellectual exercise. I have run the Clampi program only as far as needed for extracting its executable code from the layers of obfuscation, encryption and compression. Though I document the program’s Internet communications, I have not observed them at all. Though I document the program’s interface with the downloadable modules that are obviously where this malware’s threat is coded, I have not observed the interface in actual use. Yet for both those points of documentation, I expect this study will be invaluable to anyone who is observing the work of those modules.

Talking of the downloadable modules brings me to this study’s most noticeable limit. Its scope stops at the general support for those modules. I do not document any particular modules. I have not even obtained any of them for inspection, because I have drawn a line at how much unpaid work I will do for my curiosity or distraction. Regular readers of this website will be familiar with my long history of research into Windows and of writing up my results for what I see as a public good. However, I cannot justify extending this to studying malware. There is a computer security industry. Anti-virus companies have products to sell. They make money, even serious amounts of it. They employ people as analysts and reverse engineers. This self-funded study can’t hope to be as broad, but it is far more detailed than anything published by a computer security company, for probably nothing like the same number of man-hours, and I invite you to wonder how that can be.


That detail is on separate pages, one of which is quite long:

It is well known that the initial program, roughly 610KB for the copy I have studied, executes first as an installer of a persistent program. The installer extracts the persistent program, sets it up in the registry, starts it executing, and then deletes itself. The setting up arranges that the extracted program should run automatically each time Windows starts, and also provides the program with initial values for some configurable parameters. These include most notably a list of controllers to contact over the Internet and a 2048-bit RSA public key for the secure encryption of a randomly generated 448-bit Blowfish key that will protect most subsequent communication in both directions.

The persistent program, roughly 501KB for the copy I have studied, establishes secure communications with a controller and then enters a loop in which it receives commands and sends responses. Among other things, the commands provide for receiving more code to execute. The program is very much just a kernel for an extensible system of named modules. Additional modules are DLLs that are received over the Internet at the whim of the controller. They add functionality to the Clampi process and are presumably the actual agents of theft or damage. They can be executed in memory without being written to disk, or they can be saved (encrypted) in the registry and reloaded, again at the whim of the controller. Another of the commands provides for receiving a program, writing it to disk as a temporary file and running it as a new process. In short, to let Clampi run on your computer (while connected to the Internet) is to let someone unknown run whatever they want on your computer whenever they want.

Reverse Engineering

I have noted already that most reports on Clampi complain of how it is coded to frustrate inspection. The installer and kernel both have their code obfuscated so that execution first passes through hundreds of fragments for doing nonsense arithmetic before getting to anything useful. This layer of disguise is pretty much what I expect of any malware. To work through it by hand in a disassembly is a chore that you’d much prefer be done by a research student or intern. I’d have trusted it to my (admittedly bright) teenage nephews for pocket money, except that they’d have too soon got bored. Anyone who can write that this is “difficult to bypass” (see Ilomo Botnet) really should find other work (or get their employer to invest more in training).

That said, Clampi does have some obfuscation that is not so easily worked around. The kernel is written as a C++ program much like any other but it has been passed through some tool so that the x86 instructions are repackaged as bytes from the instruction set of a (simple) virtual machine. The kernel is essentially a p-code interpreter. The effect of an x86 instruction is reproduced by interpreting multiple p-code instructions, sometimes several dozen of them, and there are typically many different p-code sequences that can produce the same effect. Still, reversing the conversion from x86 to p-code looks to be a well-defined problem. I did it by hand for a few days, partly because I always value getting immersed in the data and partly to estimate the success of the obfuscation. At a rough guess, having to “recover” the x86 code before I could study it my usual way slowed me down by two orders of magnitude. That’s enough to make it impractical, of course, and the authors of this obfuscation should be pleased. On the other hand, it’s also enough to make an investment in automation seem relatively small. I, who had never much considered the problem of decompilation except to dismiss it, set about writing a tool to recover something like the original x86 code. Two weeks later I had a decompiler that was good enough to leave me with acceptably little to finish by hand. Obviously, the depth and detail that is presented in this study’s accompanying pages would never be discovered without some such decompiler.

Several reports identify the obfuscation tool and although I haven’t verified their identification for myself, I have no reason to doubt them. A page of details at the website for the cited tool, VMProtect, is certainly consistent with the obfuscation in Clampi. The tool is believably intended to help the writers of legitimate software protect their intellectual property. The expense of undoing it is much greater than the typical cost of any one copy of a program that is protected this way. Whether it is prohibitive for pirates, who make money from selling many copies of a cracked program, is another matter. To VMProtect’s customers who hope to stop their software from being reverse engineered back to its source code so that trade secrets might be built into a competing product, I say that if your competitors engage someone like me—not me, of course—to steal your work this way, then you should know that all VMProtect buys you is a few weeks of delay.

To the makers of VMProtect, I say: cooperate meaningfully with the computer security industry to ensure that your product is not used to disguise malware. Otherwise, one or another anti-virus company, or a consortium, will eventually find it worthwhile to pay someone like me—and it actually could be me—to perfect a tool that reverses yours. Though you would always be one step ahead when developing new obfuscations, being tied up with always doing that won’t be what you counted on for your business model.

To the computer security industry, I say: lift your game. Though we all know that working out what’s happening in a computer program isn’t anywhere near as easy as suggested in the movies, we reasonably expect that the computer security industry is as good at this work as anyone can possibly be. You encourage this by talking of “in-depth analysis” (see, for instance, Why is Clampi/Ilomo so effective? An analysis with detection/removal info) and of “details” in your reports, yet the reality is that you barely scratch the surface and many of your inferences from observation are incorrect on the sort of small points that count as details. Of all the reports on Clampi that I have seen for appraising my own study, only one claims to have developed a tool for reversing the work of VMProtect (see Inside the Jaws of Trojan.Clampi), and even it doesn’t (yet) have anything to show for its effort: the report contains no details that are obviously derived from its analysts having studied an approximate reconstitution of Clampi’s original x86 code, and many things that would easily be known from access to such a reconstitution are left alone or described as “obscure”. If VMProtect is at all a significant element in your collectively not having long ago worked out all that this malware can do, then either invest more in reverse engineering or stop making out that you’ve been defeated by something special.

In defence of the computer security industry, I had better add that they mostly don’t need to discover everything that any malware is capable of. The immediate commercial need is satisfied just by identifying the software as malicious and by devising reliable means of detection and removal. Clampi is not special in this sense. Though one writer (see Clamping Down on the ‘Clampi’ Trojan) says that finding it “on a computer inside your network is a little like spotting a single termite crawling into a crack in the wall”, it is in fact easy to notice and easy to remove. It can only be good that so much effort evidently has been directed at studying Clampi to understand it better than needed just to detect it and remove it. For if it is not soon removed, then according to all recent reports it can cause great damage by collecting sensitive data and sending it to who knows whom, to be acted on who knows when.

Clampi has apparently been around for years—from as long ago as 2005 according to some reports (and substantiated by details in the code). In much of that time, many have rated it as a low threat and some still do. It seems to have succeeded very well as a slow burner. If the model becomes popular, it will be important that such malware can quickly be understood more than is needed just for detection and removal. Is the industry up to the job?


The immediate practical difficulty for someone outside the industry to study any particular malware is to get a copy for inspection. Joe Stewart, of SecureWorks, may have thought my request a little odd, but without his prompt assistance this study might never have got started. I especially thank him for helpful discussions when I wondered how much to write up and which technical details might matter the most to other researchers with other methods (which are mostly a mystery to me).

Research History

I began the analysis on 23rd September. To keep my options for reporting as open as possible, I worked in a slightly unusual way, going so far as to generate a source-code approximation for every routine, no matter how insignificant. That’s a little slower than is really necessary just for understanding what a program does, but I imagined it might save time by sparing me a write-up. By 26th September, I had completed this process for the installer, and moved on to the kernel. After a few days of evaluating the VMProtect obfuscation and practising my decoding of it by hand, I decided to write a decompiler. This occupied me from 2nd October to 12th October, by which time a test run suggested to me that I had a balance between time spent on refining my automation and time spent on fixing incompleteness by hand. After a couple of days of massaging, my run from 12th October became my effective listing of the kernel’s x86 code. I then set about my usual methods of studying such listings, though again taking the unusual step of reconstituting a comprehensive approximation to source code. This was complete but for touch-ups by 2nd November and I moved on to writing a plain-language description of everything the code can do. Writing up is in some ways harder than studying the code, not least because of the discipline it imposes on you to synthesise what you can of a coherent design for the code from its many small capabilities. I considered the write-up fit to show without caution by 13th November and essentially complete by 27th November.