Threat Actors Migrating to the Cloud

April 10, 2020

Where do malware payloads come from? It’s a question with an apparently trivial answer. Usually these sit on dedicated servers owned by the campaign managers, and occasionally on a legitimate website that has been broken into and commandeered. But, as we were recently reminded, there is a third option: keeping payloads at accounts on cloud services such as Dropbox and Google Drive.

Recently, while researching the Legion Loader malware, we came across a downloader stub that, in broad strokes, did just that: downloaded a malicious payload from a well-known cloud service, and then executed it. We went looking for other similar samples, expecting to find a bounty of Legion Loaders, but the results took us by surprise. We clicked “search” and it rained: 8,000 URLs, 10,000 samples, Nanocore, Lokibot, Remcos, Pony Stealer – in short, a proper roll call of who’s who in the malware business. This isn’t one specific actor getting clever with their one specific malware. It’s a brave new service aiming to replace packers and crypters, a new fashion which cybercriminals the world over are trying on for size.

What’s in it for them? For a start, a human looking at a traffic capture generated by some software can often quickly tell whether that software is malicious or benign. Unfortunately, one of the easiest ways to tell is by looking at the domains being contacted and the contents of the transmission, which means that if some software’s entire network activity is just contacting Google Drive, a human will probably dismiss that activity as legitimate.

You might think “Ha! My firewall is not a human. Checkmate, cybercriminals”, but this sort of thing can be the difference between a working AV signature distributed in a day and a working AV signature distributed in a week. After all, researcher attention comes first and AV signatures only later. Also, your firewall probably employs some heuristic that, on its best day, emulates a human decision-maker. In that case, you’d be right to worry about malware evasion tactics that are even good enough to fool actual humans.

Google, understandably, has a zero-tolerance policy for shenanigans of this type. If you try to download malware from Google Drive, you are typically presented with the following message:

Threat Actors Migrating to the Cloud

This is one of the reasons why this “malware on the cloud” gambit didn’t just sweep the market and dominate a long time ago. If you’re looking for a web host to take your money and look the other way while you do illicit business with their web hosting, then Google is a poor choice, and you will probably have better luck with some obscure server farm in Kazakhstan. Alas, this layer of natural deterrence only goes so far, and we’ll soon see why.

When looking at recent campaigns that implement this “load malware from the cloud” method, what we’ve typically seen is spam emails that have an embedded attachment – an .ISO file that contains a malicious executable. It’s a nifty trick, but don’t bet on it staying with us for too long. Security solutions will soon learn to suspect .ISO files and inspect them thoroughly, if they didn’t before. On top of that, a human looking at such an attachment will go “huh”. Malicious campaign managers do not want victims to go “huh”. It is bad for business.

So no, the real story here isn’t the .ISO. The real story begins with the victim double-clicking the ISO and running the bundled executable. This is a stub that does not contain any functionality; instead, it downloads the malware from, say, Google Drive, and then executes it. The payload is sometimes disguised and made to superficially resemble a picture in a popular image format. So far, this is just standard downloader behavior; the stinger is that in the cloud storage, the files are encrypted. They are only decrypted on the victim machine, using “rotating XOR” decryption and a rather long key, which ranges from 200 to 1000 bytes in length and is hardcoded in the downloader stub.

This is fundamentally different from “packing” or “crypting” malware. Packed malware appears to be gibberish, but will reveal its function and behavior during execution; encrypted malware stays gibberish as long as you don’t have the key. Now, in theory, it so happens that rotating-XOR encryption with a 1000-byte key can in fact be broken if the plaintext is long enough and coherent enough (there’s even a tool that will do it for English plaintexts). In practice, we don’t live in a world where defenders launch cryptographic attacks at suspicious binary blobs just in case something interesting turns up. The performance overhead is just too prohibitive.

Worse: even if Google were determined to force these encrypted payloads to reveal their secrets, malicious actors could then route the attack permanently without too much effort. We’ll leave out the details, as we are loath to give malware authors ideas, and consider the typical cryptographic illiteracy in the cybercriminal crowd as a gift that should be handled with great care. Suffice it to say that if you know your crypto 101 then you know that these payloads could be processed such that even if Google throws their fanciest GPU rigs and their cleverest algorithms at the problem, these will bounce right off.

This is a right mess. What do these “terms of service” even mean if they cannot be directly enforced? We sympathize with Google, who cannot do much more than employ the stop-gap measure of looking for plain malicious binaries and praying that this practice doesn’t catch on. Of course they can also follow the payloads when campaigns come to light, investigate the uploads, follow the leads, create deterrence. But this is complicated, manual, delayed. Cybercriminals love to force their security-minded adversaries into complicated, manual and delayed responses. It’s what they live for.

So the cloud host doesn’t kick the malicious payload off their servers, because they can’t, and the user runs the dropper and the dropper fetches its payload. An image is also displayed to the user, presumably to cover the attack’s tracks a bit. Can’t be malware if you actually got to see your very urgent Jury Duty summons as promised in the spam message, right? Right.

Decoy image displayed to the user

Figure 1 – Decoy image displayed to the user.

Each payload is encrypted using a unique encryption key (the bad guys got this part right, sadly). The dropper also has a built-in option in its hardcoded configuration that allows for deferred downloading of the payload after system reboot.

Figure 2 – Infection flow involving the dropper.

To add insult to injury, the malicious payload is stored only in memory and is never saved to the disk in either decrypted or encrypted form. We’d call this “fileless”, except, you know, the original dropper is a file. We suppose we could say that it is fileless with respect to the decrypted payload. The threat model in the attackers’ minds is very clear: Google and security vendors are all looking at files, looking for familiar signatures and hashes. Never put the malicious fully-formed binary in a file and, as an attacker, you’re home free.

Does this model reflect reality? Well, yes and no. Yes, some victims will, regrettably, have about this level of security. No, this isn’t the limit of what security solutions can actually do in this case, and hasn’t been for nearly fifteen years. For instance, a sandbox environment will emulate the dropper’s execution – complete with the malicious payload being downloaded and executed, and the resulting incriminating behavior. The sandbox will then deliver a verdict: “you probably shouldn’t run this file on your own machine”.

But if a sandbox doesn’t record the whole interaction as it happens, defenders don’t have much recourse after the fact. When a campaign ends, the encrypted malicious sample is removed from the cloud storage – leaving researchers to look at a featureless stub downloader, an encryption key and a dead cloud storage link. Typically, no traces will remain on the victims’ machines to investigate the data leak, either. The malicious binary only existed in volatile memory, and by the time an analyst gets to look at the machine, the offending code has long since scattered to the four winds.

GuLoader distributed malware

It’s worth to mention that in a small number of the cases we examined, these encrypted malicious payloads were hosted at compromised legitimate websites. Encrypting the payload in such a scenario is probably overkill (we can’t help but visualize the threat being staved off here, which is the owner of putting down her tea cup, straightening her glasses and pouting, “now wait here just a moment, what’s this binary I am seeing in my directory listing? Lord be my witness, I’ll be uploading it to that newfangled VirusTotal website right this second, and that file better pray it doesn’t come back as one of those Emotets or Nanocores. These hacker brats! The nerve!”).

Figure 3 – Cloud services used for downloading payloads

Figure 3 – Cloud services used for downloading payloads.

It’s also worth to mention that Google Drive and OneDrive weren’t the only unwitting carriers of encrypting payloads. Some other services were used, even if sparingly:

Service Number of samples 48 30 26 17 14 13 10

We still continue to see approximately 800 new samples of this dropper per week.

Analysis Story and Technical Details

Now that you understand why this downloader is such a nuisance, we bet you want a look at the nuts and bolts of the research, and we’re happy to indulge you.

As mentioned, this story began with a campaign delivering the Legion Loader malware. We noticed that VirusTotal behavior analysis reports for this malware family contained DNS requests to The analyzed samples were very small and couldn’t possibly contain the researched malware even in packed form. It was obvious that the analyzed samples were just droppers capable of downloading and executing the malware.

Figure 4 – Dropper execution flow

Figure 4 – Dropper execution flow.

Shellcode decryptor

The dropper (hiding in the ISO file, remember) is crafted to appear, at first sight, as a Visual Basic 6 application. It usually has very few functions recognized by disassemblers, and even those contain a lot of junk instructions mixed in with obfuscated code, anti-disassembly measures and plain noise for a general holistic experience of analyst misery.

Figure 5 - Jump obfuscation in the dropper

Figure 5 – Jump obfuscation in the dropper.

Figure 6 - IDA fails to analyze the code

Figure 6 – IDA fails to analyze the code.

As a result, there is much manual grunt-work of figuring out where the execution flow goes and forcing the disassembler to interpret the instructions there as code. The end result is still an eyesore, but some careful analysis will show that it decrypts and executes the shellcode, which is located somewhere else in the binary (typically in the resources or in the code section).

Figure 7 - Calculation of the shellcode decryption key

Figure 7 – Calculation of the shellcode decryption key.

Don’t mistake this decryption with the decryption we were complaining about earlier. Yes, it’s also a rotating XOR decryption, but the key is much shorter – and, more importantly, it is right there. “Encryption but the key is right there” is just a long-winded way to say “obfuscation”, and it is not nearly as much of a headache. The dropper elegantly recovers the 4-byte key by XORing the correct first 4 bytes of the plaintext with the first 4 bytes of the ciphertext:

plaintext_prefix = 0x0200EC81;
key = ((DWORD *)ciphertext)[0] ^ plaintext_prefix;


The shellcode is also obfuscated (because of course it is), making IDA unable to automatically analyze it. It also contains some anti-debugging tricks, in case you were thinking to throw it in a debugger.

Figure 8 - Anti-debug trick used in the shellcode

Figure 8 – Anti-debug trick used in the shellcode.

For example, in the code above the malware hides the current thread from the debugger. This leads the debugged application to crash on any breakpoint hit in the hidden thread.

The dropper prevents the debugger from attaching to the running process by hooking the DbgUiRemoteBreakin function,  and redirecting execution to an invalid address pointed by uninitialized variable (likely by mistake, because this kind of anti-debug technique typically uses redirection to ExitProcess).

It also replaces the DbgBreakPoint function body with a NOP operation. This prevents a debugger from breaking when it attaches to the process. DbgUiRemoteBreakin and DbgBreakPoint are the key functions that are called when a debugger attaches to a process; by meddling with them, the shellcode cripples the ability of a naïve analyst to debug the process. If you’re interested in a more detailed take on how these anti-debugging techniques work, we recommend this github repository.

Figure 9 - Anti-debug trick used in the dropper

Figure 9 – Anti-debug trick used in the dropper.

Remember that we earlier sang the praises of sandboxes as a solution to this sort of gambit by cybercriminals; sadly, they know well enough to worry about sandboxes, which is why the shellcode also includes a host of techniques to check if it’s running in a sandbox, and refuse to run if the answer is positive. These are called “evasions”, and there are many (many) of them. The researched sample in particular checked the number of top level windows; If this number is less than 12, the dropper silently exits. (You can read more about this evasion at this entry in our Sandbox Evasion encyclopedia).

Figure 10 - Sandbox evasion technique used in the dropper

Figure 10 – Sandbox evasion technique used in the dropper.

The dropper dynamically resolves API functions, which has been par for the course for a long while now (analysts still check the imports when looking at a new malware, but they know full well that this is just a formality). Function names are stored in the code right after calls to procedures that have no returns.

Figure 11 - Resolving API function addresses

Figure 11 – Resolving API function addresses.

After resolving the addresses of the API functions, the dropper launches another process of itself in a suspended state. The malware unmaps its image from the image base of this child process, maps the msvbvm60.dll library there (at 0x400000), allocates memory in the child process, copies the decrypted shellcode into the allocated memory, and transfers execution there.

Figure 12 – Copying the shellcode into child process memory

Figure 12 – Copying the shellcode into child process memory.

The parent process at this point has served its purpose, and is terminated.

Downloading payload

The shellcode downloads the encrypted payload from the hard-coded URL.  In 72 % of samples is used for downloading payloads:

Figure 13 - URL for downloading malicious payload found in the shellcode

Figure 13URL for downloading malicious payload found in the shellcode.

The payload is downloaded using functions InternetOpenUrlA and InternetReadFile:

Figure 14 - Payload downloading routine

Figure 14 – Payload downloading routine.

Using the following hardcoded user-agent:

Mozilla/5.0 (Windows NT 6.1; WOW64; Trident/7.0; rv:11.0) like Gecko

The only way the dropper checks the consistency of the downloaded payload is by comparing its size with a hard-coded value. The dropper repeatedly tries to download the payload in an infinite loop until the downloaded file size is equal to the expected value.

In most cases the same pattern is used for naming encrypted files stored in cloud drives:

<prefix>_encrypted_<7 hex digits>.bin

Figure 15 – Encrypted file name

Figure 15 – Encrypted file name.

Encrypted payload starts from 64 hex digits sequence, which is not used in the decryption process:

Figure 16 – Encrypted payload part

Figure 16 – Encrypted payload part.

As we explained earliest, to recover a working binary from the downloaded data, the dropper uses XOR operation with a unique key that’s typically several hundred bytes in length.

Figure 17 - A part of the payload decryption key

Figure 17 – A part of the payload decryption key.

The payload decryption routine is obfuscated as well:

Figure 18 - A part of the payload decryption routine

Figure 18 – A part of the payload decryption routine.

This is equivalent to a rotating XOR decrypt, written below in Python for your convenience:
decrypted_data = [data[i] ^ key[i % len(key)] for i in range(len(data))]
The decrypted payload is manually loaded to its image base address that is extracted from the PE header of the payload.

The dropper then creates a new thread to run the payload without creating a separate process. Next, the malware hides the thread, in which payload is executed, and terminates the other thread that did all the decryption work.

Figure 19 - Hiding the main thread of the payload

Figure 19 – Hiding the main thread of the payload.

Decoy images

We also observed samples containing two URLs. The second URL is used for downloading the decoy image that will be displayed to the user.

Figure 20 - URL for downloading decoy image

Figure 20 – URL for downloading decoy image.

The downloaded image is saved to user profile folder under the hardcoded name. Then the image is displayed using the ShellExecuteW API function.

The cybercriminals currently use a limited set of images. Here are some of them:

Figure 21 – Decoy images

Figure 21 – Decoy images.

In different samples we found the following URLs used for downloading images:

URL Filename perez.jpg Manuela7.jpg Manuela1.jpg 60.jpg mark.jpg marek72.jpg MN1.jpg as1.jpg mr.jpg Manuela5.jpg Manuela6.jpg Manuela82630.jpg

Table 1 – URLs for downloading decoy images.

Delayed downloading

Depending on the dropper’s configuration, it is capable of setting up deferred download of the payload after reboot.

If this option is enabled, the dropper reaches for the registry’s autorun key (a humble technique as old as time – some analysts can recite the entire key path in their sleep):

Figure 22 - Registry autorun entry

Figure 22 – Registry autorun entry.

The dropper is copied to %USERPROFILE%\subfolder1\filename1.exe and a small VBS script (C:\{USERPROFILEPATH}\subfolder1\filename1.vbs) is created with the following content:

Set W = CreateObject("WScript.Shell")
Set C = W.Exec ("C:\Users\User\subfolder1\filename1.exe")

With “User” being the appropriate username. After the OS reboots, the dropper is started from the %USERPROFILE%\subfolder1\filename1.exe. Only then it downloads and execute the payload. All of this, of course, assuming this “deferred download” option was turned on.

We’re not Robots, We Swear

While trying to download and analyze the malicious samples from the cloud drives we faced an issue. After downloading several of these files from different Google Drive URLs, Google banned our IP address and we started getting the following error message:

your computer or network may be senging automated queires

We were thus faced with a situation in which Google protection against bots made our work more difficult, blocking us from downloading malicious samples. However, as the dropper downloads only one file from a cloud drive, it is not affected by this issue.


The term “perfect storm” should be used very sparingly, so let us say that this dropper is a Strong Breeze of troublesome features. The burden of server maintenance is shifted to Microsoft and Google, who will have trouble being proactive about the issue. Many (not all) popular and formidable-sounding security solutions are just inadequate in the face of this threat. The malware leaves no trace once it is done running, and even if an analyst gets their hand on a copy they then have to contend with a host of anti-analysis measures (and Google Drive’s security features to top it off).

Having said all that, there is also some good news.

First of all, the science of cybercrime marches slowly. Progress is forgotten, worst practices endure and multiply. If we see this clever alternative to packing wallow in obscurity and die out in favor of the 294th API-bombing packer and the 1077th transmogrified UPX, it won’t be the first promising advance in malware-faring that we’d seen gone and forgotten, or the last. (Do you remember ransomware encrypting offline using Diffie-Hellman Key Exchange with the criminals’ master key? What happened to that?)

Second of all, Sandbox Analysis makes short work of the entire thing, as long as evasions are pre-emptively dealt with (and there’s no reason why they shouldn’t be). Like many apparent next-level threats, defusing this dropper doesn’t require some super advanced technology on the bleeding edge – just good fundamentals, or in this case, judicious use of technology that’s been around since before the original iPhone.

We often worry about the day when cybercriminals finally understand how easily they could make all our lives really difficult, but happily, today is not that day.



MD5 Embedded URLs





  • Check Point Research Publications
  • Global Cyber Attack Reports
  • Threat Research
February 17, 2020

“The Turkish Rat” Evolved Adwind in a Massive Ongoing Phishing Campaign

  • Check Point Research Publications
August 11, 2017

“The Next WannaCry” Vulnerability is Here

  • Check Point Research Publications
January 11, 2018

‘RubyMiner’ Cryptominer Affects 30% of WW Networks