Research By: Nadav Grossman
Since the revelation of the EternalBlue exploit, allegedly developed by the NSA, and the malicious uses that followed with WannaCry, it went under thorough scrutiny by the security community. While many details were researched and published, several remained in the dark, and an end-to-end explanation of the vulnerability and exploit is nowhere to be found.
For these reasons, we endeavored to conduct a comprehensive research of the vulnerabilities and exploitation related to EternalBlue, by using reverse engineering, kernel debugging and network analysis. The results of this research appear below. We hope that this paper will enhance the understanding of EternalBlue, and help the security community combat it, and the likes of it in the future.
The EternalBlue exploitation tool was leaked by “The Shadow Brokers” group on April 14, 2017, in their fifth leak, “Lost in Translation.” The leak included many exploitation tools like EternalBlue that are based on multiple vulnerabilities in the Windows implementation of SMB protocol.
EternalBlue works on all Windows versions prior to Windows 8. These versions contain an interprocess communication share (IPC$) that allows a null session. This means that the connection is established via anonymous login and null session is allowed by default. Null session allows the client to send different commands to the server.
The NSA created a framework (much like Metasploit) named FuzzBunch, which was part of the leak. The purpose of this framework is to configure, for example, set victim
ip, and execute the exploitation tools.
Microsoft released patches for the vulnerabilities in the leak, under the MS17-010 (Microsoft Security Bulletin).
CVE-2017-0144 is the CVE ID in MS17-010 that is related to EternalBlue.
Server Message Block (SMB), one version of which is also known as Common Internet File System (CIFS), operates as an application-layer network protocol mainly used for providing shared access to files, printers, and serial ports and miscellaneous communications between nodes on a network.
It also provides an authenticated inter-process communication mechanism. Most SMB use involves computers running Microsoft Windows, where it was known as the “Microsoft Windows Network” before the subsequent introduction of Active Directory. Corresponding Windows services are LAN Manager Server (for the server component) and LAN Manager Workstation (for the client component) [Wikipedia].
EternalBlue exploits 3 bugs (named as Bug [A,B,C]) to achive RCE, the explanation for each bug are listed below:
A bug in the process of converting FEA (File Extended Attributes) from Os2 structure to NT structure by the Windows SMB implementation (srv.sys driver) leads to buffer overflow in the non-paged kernel pool.
A bit about FEA (File Extended Attributes):
The FEA structure is used to describe file characteristics. For more information, see https://en.wikipedia.org/wiki/Extended_file_attributes.
You can think about the FEA structure as key and value pairs, where the key is
AttributeName and the value is
In Os2 format, the struct looks like this:
After you convert the FEA from Os2 format to Windows format (Nt Format), the structure looks like this:
Note– These structures are not valid
Structs, because the length of the AttributeName and AttributeValue changes (is not a constant). However, it’s simpler to understand the structures using this illustration.
The functions listed below are part of the srv.sys driver and they are related to the bug.
SrvOs2FeaListToNt – Converts Os2 FEA List to NT FEA List.
The logics of this function are:
SrvOs2FeaListSizeToNTto get the appropriate size for
Os2FeaListuntil it reaches the
Os2FeaList). In each iteration, it calls
SrvOs2FeaToNTto convert the
Os2Fearecord to NT format and add it to
SrvOs2FeaListSizeToNT – Calculates the size needed to convert
Os2FeaList structures into the appropriate
The logics of this function are:
NtFeaList. The buffer size is calculated according to the required size to convert
NtFeaListstruct (this is the return value from the function).
Os2FeaListshould be converted to
NtFeato store them later (by
NtFeaList. The result of the calculation is stored in the member of
SizeOfListInBytes, by overwriting the previous value .
SizeOfListInBytesand the remainder of the FEA is “out of range” (bigger than
SizeOfListInBytes). It will ignore this FEA and any further “out of range” FEA and shrink the
SizeOfListInByteto the size of all the “valid” FEAs.
SizeOfListInBytesdoesn’t restrict the SMB packet size.
SrvOs2FeaToNT – Converts the
Os2Fea record to an
NtFea record. The structure is not identical to the
NtFea, but it’s quite similar.
*Note that the size of the
NtFea record is bigger than
Os2Fea because it contains another field named
NextEntryOffset. There is also an alignment of 4 bytes between the
SrvOs2FeaListSizeToNT shrinking illustration:
After Shrinking: if the size of
SizeOfListInBytes is below 216:
After Shrinking (bug): if the size of
SizeOfListInBytes is above 216:
In a nutshell: There is a wrong casting in the
SrvOs2FeaListSizeToNT function, in its second logic (shrinking the
SizeOfListInBytes member of
Os2FeaList). That leads to a small buffer (i.e.
NtFeaList) and large data (out of buffer boundary) that will be stored on it after conversion (i.e.
Instead of shrinking the
SizeOfListInBytes, the function enlarges it.
SizeOfListInBytes indicates how many records (in bytes) of
Os2Fea should be converted to
However, the size of the
NtFeaList buffer that returns from the function (
NtFeaListSize) is calculated correctly for the appropriate size to convert a shrunk
This leads to a situation in which the size of bytes that should be copied to the buffer (
NtFeaList, which depends on the updated value
SizeOfListInBytes member of
Os2FeaLis) is bigger than the buffer size (
NtFeaListSize). This causes an Out of Bound write.
SizeOfListInBytes is a DWORD size member of
Os2FeaList. If shrinking is needed, the function treats
SizeOfListInBytes as a Word member and it updates only the 2 bytes, instead of 4 bytes (with the shrunk size). The 2 most significant bytes remain untouched. This bug causes
SizeOfListInBytes to be enlarged instead of shrunk.
If the value of
SizeOfListInBytes is in the range of 216, there is no problem and the function works as expected.
However, if the value of
SizeOfListInBytes is above the range of 216, it could be enlarged instead of shrunk.
In the 2 examples below, the first is of a normal scenario (if the bug wouldn’t have existed) and the second is of the buggy scenario.
* The important lines are marked in Yellow
When you transmit a file over SMB protocol, there are several data-related functions:
1. SMB_COM_TRANSACTION2: Sub-commands provide support for a richer set of server-side file system semantics. The “
Trans2 subcommands”, as they are called, allow clients to set and retrieve Extended Attribute key/value pairs, make use of long file names (longer than the original 8.3 file format names), and perform directory searches, among other tasks.
2. SMB_COM_NT_TRANSACT: Sub-commands extend the file system feature access offered by
SMB_COM_TRANSACTION2, and also allow for the transfer of very large parameter and data blocks.
If the data sent via
SMB_COM_TRANSACTION2 or by
SMB_COM_NT_TRANSACT exceeds the
MaxBufferSize established during session setup, or
total_data_to_send is bigger than
transmitted_data, then the transaction uses the
Each sub-command has a corresponding sub-command
_SECONDARY is used when the data sent is too big for a single packet. It is therefore split over a few packets to fulfill “the total size of data to be sent” that was declared in the first packet. The packets that follow the first sub-command have the corresponding
_SECONDARY sub-command set as their command.
SMB_COM_NT_TRANSACT => SMB_COM_NT_TRANSACT_SECONDARY
SMB_COM_TRANSACTION2 => SMB_COM_TRANSACTION2_SECONDARY
SMB_COM_TRANSACTION2, the maximum data that can be sent is represented by a parameter in the header of
SMB_COM_TRANSACTION2 in the field of a
Word size. The same is true for the
SMB_COM_NT_TRANSACT, the maximum data that can be sent is represented by a parameter in the header of
SMB_COM_NT_TRANSACT in the field of
Dword size. The same is true for the
Therefore, there is a difference between the amounts of data that can be sent in
SMB_COM_TRANSACTION2, where the maximum data length is represented in a
0xFFFF), and in
SMB_COM_NT_TRANSACT where the maximum is represented in a
However, as there is no validation for which function started the transaction (
SMB_COM_NT_TRANSACT), parsing is according to the last transaction type.
Thus, it’s possible to send:
SMB_COM_NT_TRANSACT followed by
This situation can lead to wrong data parsing, and this bug enables
Bug A by treating
It happens when
Dword) is followed by
Word). This leads to parsing the data, wrongly, as if it originally came from a transaction of the
In a nutshell: There is a bug that lets you allocate a chunk with a specified size in the kernel non-paged pool with the specified size. It is used in the heap grooming phase when creating a hole that later will be filled with a data size that causes an out of bound write to the next chunk (
Bug A &
SMB_COM_SESSION_SETUP_ANDX request MUST be sent by a client to begin user authentication on an SMB connection and establish an SMB session.
This command is used to configure an SMB session. At least one
SMB_COM_SESSION_SETUP_ANDX MUST be sent to perform a user logon to the server and to establish a valid UID.
There are 2 formats for an
The first is used for
NTLM authentication, documented here. This is the format of the request:
The second is used for
NTLM SSP) authentication, documented here. This is the format of the request:
In both formats, the request is split into 2 sections:
SMB_Parameters– Contains parameters of sizes between 1-4 bytes. The
WordCountfield represents the total length of
SMB_Parametersstruct members in a
SMB_Data– Contains data in a variable size. The
ByteCountfield represents the length of the
SMB_Datastruct members section in bytes.
Summing the size of the fields, in the first format, the
WordCount equals 13 and in the second format (
extended security), the
WordCount equals 12.
The server does an integrity check, with a function named
SrvValidateSmb, for SMB packets that include the format of
Although there is no bug in the
SrvValidateSmb function, there is a bug in the extraction of
SMB_DATA by a function named
By triggering the bug, the
BlockingSessionSetupAndX function wrongly calculates
ByteCount, which leads to an allocation of controlled size – bigger than the packet data – in the non-paged pool.
Here is the buggy part:
SMB_COM_SESSION_SETUP_ANDX request is handled by the
BlockingSessionSetupAndX function. Below is pseudo code for handling both request formats (only the relevant part is shown).
From the pseudo code above, we see that if we send an
SMB_COM_SESSION_SETUP_ANDX request as
Extended Security (
WordCount 12) with
CAP_EXTENDED_SECURITY, but without
FLAGS2_EXTENDED_SECURITY, the request will be processed wrongly as an NT Security request (
WordCount 13, marked in yellow).
In this case, the
GetNtSecurityParameters function is called, but it calculates the
SMB_DATA wrongly. The request is in
Extended Security (
WordCount 12) format, but the function intends to parse it as
NT Security request (
As a result, it reads
ByteCount from the wrong offset in the struct, and allocates space in the non-paged kernel pool for
NativeLanMan unicode strings (see the structure above), according to the wrong offset of ByteCount.
This bug allows you to send a small packet that leads to a big allocation in the non-paged pool, which is used to create a big allocation as a placeholder.
This allocation will later be freed (creating a
HOLE) and allocated again by an
NtFea chunk that will overflow the next chunk. This issue is explained further in the exploitation flow section.
* See the glossary section for an explanation of terms.
pMdl1) Overwrite – When we get a write primitive to an
MDL, we can map an I/O data to a specific virtual address. In the exploit we have an OOB (out-of-bound) write in
Bug A and Bug B). We can therefore overwrite the header of
srvnetchunk (using some sort of grooming).
srvnetchunk there is a
MDLAs explained in the glossary section, a virtual address in the MDL maps the incoming data from the client to a specific virtual address. Therefore, if we change it by some sort of overwriting, the data from the client is mapped to the address that would have been overwritten. It enables the client data (that we control) to be written wherever we want (controlling the
pSrvNetWskStructis located in the
srvnetheader struct and points to the
SrvNetWskStruct. This struct contains a pointer to the function
HandlerFunction, which is called when the related
srvnetconnection is closed. If we overwrite the pointer to the
pSrvNetWskStruct) with an address that contains a fake
SrvNetWskStructwhose values we control, we can control the value of the pointer to the function that is called when the
srvnetconnection is closed (
HandlerFunction), which leads to RCE.
MDL(primitive 1), we can control where the data under our control (data sent from the client to the server, over a connection that is related to
srvnet) would be written.
StartVAfield) of the
MDLa static address like an address within
HAL's Heap, the data that the client sends will be mapped to a virtual address within the
HAL's Heap. (
HAL's Heaphas execute permission in Windows versions prior to Windows 8).
HAL's Heap(by primitive 1) we must combine it with primitive 2.
pSrvNetWskStruct) to point to the same address that was crafted in the
MDL(static address on the
HAL's Heap), in that address (
HAL's Heap) we can create a fake struct (
SrvNetWskStruct) preceded by the shellcode. The shellcode would then be called upon closing the related srvnet connection, and we achieve RCE.
SrvNetWskStruct struct contains a pointer to a function (
HandlerFunction) that is called when
srvnet connection is closed. The pointer to function (
HandlerFunction) should point to our shellcode.
Assume that this is the initial state of the non-paged kernel pool and the
srv allocation according to
Bug A and Bug B. In this step, only the connection for
Os2Fea transmission is opened.
Fills part of the
OS2Fea first by
SMB_COM_NT_TRANSACT. Later, it is filled by
SMB_COM_TRANSACTION2_SECONDARY as described in
Bug A and
Bug B, but without sending the last
SECONDARY packet (this does not yet trigger the OOB write).
srvnet connections to increase the chances of
srvnet overwriting (overflow) by preceding the
srv allocation of converted
Bug A and
Bug B, this leads to overflowing the next chunk (
srvnet Header). This technique is also used as a grooming technique.
At this point, the attacker creates a chunk according to
Bug C. This chunk is used as a placeholder for the converted
NtFea. It should be the same size as the overflowing
NtFeaList (wrong calculation, due to
Bug A). This chunk is later freed by closing the connection, and
NtFea (write out-of-bound) is allocated instead (fills the hole).
Before it is freed, more
srvnet chunks (new
srvnet connections) are allocated.
It is freed just before the final packet of
srv allocation (
SMB_COM_TRANSACTION2_SECONDARY) that allocates a chunk for storing the
NtFea converted data.
Most likely, this allocation is stored in this freed chunk. It’s part of the grooming technique.
srvnet allocation (by a new connection). This
srvnet chunk is located after “the chunk of
Bug C.” If the
NtFea is located in the previous chunk (“the chunk of
Bug C“), it would lead to an overflow of this
The “chunk of
Bug C” is freed.
1. The last
Os2Fea data is sent (using
SMB_COM_TRANSACTION2_SECONDARY). This leads to allocation to convert
NtFea is allocated at the free hole (previously was allocated with the chunk according to
NtFea overwrites (overflow) the next chunk, which is a
srvnet chunk. The
srvnet header is overwritten and modifies these 2 header properties to point to the same address in the
pSrvNetWskStruct, the pointer to the struct that includes the function that is called when the connection is closed.
MDLthat is used to map the next incoming data (from the user) in this
srvnetconnection from the user arrives.
MDL, this data is written to the
HAL's Heap, instead of the reserve place in the kernel pool (the value of the
MDLbefore the overwrite).
HAL's heapcontains a fake struct (
pSrvNetWskStructpoints to the fake struct (
SRVNET_RECV) in the
HAL's heap. When the connection closes, the
HandlerFunctionfrom the fake struct is called.
DoublePulsar. It is written after the fake struct to the
srvnetconnections are closed.
HandlerFunctionpointed at by the
pSrvNetWskStruct) is fake and points to the fake struct in the
HandlerFunctionis executed, but this function is the shellcode.
This is an edited wireshark pcap of a successful exploitation on Windows 7. It includes only one exploitation try; in reality, it usually takes more than a single attempt. We added relevant filters, coloring and comments. Responses were omitted for readability.
Client systems use the Common Internet File System (CIFS) Protocol to request file and print services from server systems over a network. CIFS is a stateful protocol, in which clients establish a session with a server and use that session to make a variety of requests to access files, printers, and inter-process communication (IPC) mechanisms, such as named pipes. CIFS imposes state to maintain an authentication context, cryptographic operations, file semantics such as locking, and similar features.
The Server Message Block (SMB) Version 1.0 Protocol extends the CIFS Protocol with additional security, file, and disk management support. These extensions do not alter the basic message sequencing of the CIFS Protocol but introduce new flags, extended requests and responses, and new information levels. All of these extensions follow a request/response pattern in which the client initiates all of the requests.
The Server Message Block (SMB) Protocol Versions 2 and 3 supports the sharing of file and print resources between machines. The protocol borrows and extends concepts from the Server Message Block (SMB) Version 1.0 Protocol.
These drivers are related to SMB protocols:
srv.sys– Related to
srvnet.sys– Related to
SMBv1 packets from the client to the server cause allocation in a paged or non-paged kernel pool.
Bug A and
Bug B use the
SMBv1 protocol that leads to
srv allocation. The out-of-bound write happens in
SMBv2 packets from the client to the server cause allocation in a paged or non-paged kernel pool. In the exploit, the srvnet chunk is allocated, so we can overflow it with
srv allocation. By doing this, we overwrite a structure in the srvnet header that leads to RCE (Remote Code Execution)
This struct is pointed to by the
pSrvNetWskStruct variable in the
SRVNET_HEADER. This is used by
srvnet.sys to call the handler function (which points to the shellcode address in the EternalBlue scenario) when the connection is closed.
This struct resides in the beginning of the srvnet related chunk. In the exploitation, we use/manipulate it by overwriting these fields:
pSrvNetWskStruct: Pointer to the
pMdl1: Pointer to a
MDL. This maps received data to virtual address space.
A memory descriptor list (MDL) is a system-defined structure (kernel structure) that describes a buffer by a set of physical addresses. A driver performs direct I/O receives a pointer to an MDL from the I/O manager, and reads and writes data through the MDL.
An I/O buffer that spans a range of contiguous virtual memory addresses can be spread over several physical pages, and these pages can be discontinuous. The operating system uses an MDL to describe the physical page layout for a virtual memory buffer. For more info, see references 1, 2.
MDL lets you “write-what-where” the primitive.
MDL is “locked”, specific physical pages are mapped to specific virtual address space.
For example: It is possible to map the entire virtual page range to one physical page. No copy is needed because every request (for example, read/write) to a specific virtual range will fetch the same physical page.
HAL's Heap – The Windows HAL’s heap is located at 0xffffffff’ffd00000 in 64-bit and at 0xffd00000 in 32-bit (the address was static until Windows10). This area is executable in Windows versions prior to Windows 8 / Server 2012.
In this research paper, we attempted to shed light on important sections of the EternalBlue vulnerabilities and exploit, and provide a full step-by-step explanation of the causes for the vulnerabilities and their exploitation. It is our belief that with better understanding of the roots of EternalBlue, we will be able to improve the security of users around the world.
As important as EternalBlue is, it is not the first, nor the last major exploit that enables hackers to take complete control over entire networks. Unfortunately, nefarious exploits continue to appear, as was recently shown by Microsoft itself. Users and organizations cannot wait for a second round against cyber threats, and must protect themselves with the most up-to-date protections, capable of detecting and blocking any cyber-attack on its tracks.
I would also like to thank my colleagues Omri Herscovici, Yannay Livneh, and Michael Kajiloti for their help in this research.
The Check Point IPS blade provides protections against these threats:
Microsoft Windows EternalBlue SMB Remote Code Execution