Blog / Case Study · Threat Analysis

The Key Outlived the Patch: Warlock's Path from SharePoint to SYSVOL

A year after ToolShell, the group behind Warlock ransomware still gets into on-premises SharePoint — including servers that were patched. Nineteen steps, nine days, one stolen key that the patch never touched.

Bernhard Kreinz • • Loading... • TLCTC v2.6
In one paragraph

Symantec and Carbon Black have reconstructed a July 2026 intrusion by the group they call Longlegs (also Storm-2603) at a critical-infrastructure operator: in through on-premises SharePoint, out nine days later with Warlock running on at least 33 hosts. Classified by cause it is nineteen steps, and the one that matters most is the second: the webshell stole the SharePoint farm's ASP.NET machine keys. From that moment the attacker no longer needed the vulnerability. A forged, validly signed ViewState is a credential being used — #4, not #2 — and patching the code does not revoke it. That is why the report speaks of exploitation of “both patched and unpatched” servers. The control that closes the path is key rotation, not another patch.

What Happened

Warlock first appeared in mid-2025, when its operators were among the groups exploiting the SharePoint “ToolShell” zero-days. Symantec tracks them as Longlegs; Microsoft calls them Storm-2603, Secureworks Gold Salem. In their new report the Symantec and Carbon Black Threat Hunter Team count at least four victims in the past two months — a water utility, a telecommunications provider, a regional government body and a university, all in Portuguese- and Spanish-speaking countries across Europe, Africa and Latin America. Their conclusion: “exploitation of ToolShell and other related-SharePoint vulnerabilities remains a viable initial access route for attackers SharePoint deployments that have not been patched or otherwise mitigated.”

One intrusion, at a critical-infrastructure operator, is documented down to the command line:

Date (2026)What the telemetry shows
22 JulyA PowerShell one-liner writes the webshell layout2sp.aspx into the SharePoint LAYOUTS folder of several SharePoint versions at once; its job is to harvest the farm's ASP.NET machine keys.
24–28 Julywhoami, net user /domain, nltest /domain_trusts on the SharePoint hosts.
28 JulyA forged __VIEWSTATE, signed with the stolen key, runs code in the SharePoint application pool. Three MSI packages arrive through msiexec /q /i from catbox and wasabisys within ninety minutes; signed executables side-load the attackers' DLLs.
28–29 JulyThe account SPSEPRDSetup is added to the local Administrators group on three more hosts; a Visual Studio Code tunnel is installed as a service on one of them.
30–31 JulyNetExec for Active Directory enumeration, credential spraying and remote command execution.
31 July, early morningAn AV/EDR killer using the vulnerable driver K7RKScan.sys (CVE-2025-1055) runs on at least 40 hosts within about two hours. Warlock is staged in the domain's SYSVOL share and executes on at least 33 hosts.

The headlines call this a ransomware attack. TLCTC does not classify the headline. “Ransomware” is the outcome of the last step (Axiom III); the question is which generic vulnerability each step exploited.

The Path

#2 ||[prod][@External→@Org]|| →[Δt=0s] #7 + [DRE: C] →[Δt≤6d] #4 →[Δt=0s] #7
→[Δt=~days] #1 →[Δt=~h] #1 →[Δt=~min] #7 →[Δt=~1d] #4 [inferred] →[Δt=~min] #1
→[Δt=~h] #1 ||[cloud][@External⇒@Microsoft→@Org]|| →[Δt=~1d] #7 →[Δt=~min] #1
→[Δt=?] … →[Δt=?] #4 →[Δt=s] #1 →[Δt=s] #7 →[Δt=s] #2 |[privilege][@admin→@kernel]|
→[Δt<2h] #1 →[Δt=~h] #7 + [DRE: Ac]
#StepClusterWhy
1ToolShell chain against on-prem SharePoint (CVE-2025-49706 authentication bypass, CVE-2025-49704 / -53770 deserialization, -53771)#2Implementation flaws in software acting as a server. The bypass presents no credential, so it is not #4.
2Dropper and webshell run; the farm's machine keys are read#7 + [DRE: C]R-EXEC. The keys are credential material; the DRE sits at the acquisition (R-CRED).
3A ViewState signed with the stolen key is accepted as authentic#4A stolen secret applied — credential use, separate from its theft (Axiom X).
4The System.Workflow gadget runs the attackers' commands in the app pool#7R-EXEC. With a valid MAC, deserializing the ViewState is designed behaviour — no second flaw.
5whoami, net user /domain, nltest#1Built-in commands used as designed.
6msiexec /q /i from catbox and wasabisys#1A designed install-from-URL function.
7Signed binaries side-load the attackers' DLLs#7The attackers' code executes; side-loading is how it is loaded.
8Admin session on hosts 3–5#4 [inferred]Required for step 9; the logon itself is not in the report.
9SPSEPRDSetup into local Administrators#1A designed privileged function.
10code-insiders.exe tunnel service install#1Microsoft-signed tool, designed feature; Microsoft's relay is transit (⇒).
11NetExec runs#7Attacker-supplied offensive tooling executes.
12AD enumeration and remote execution through NetExec#1Designed LDAP / SMB / remote-administration functions.
13Domain-wide credentials obtained…Unresolved: the report does not say how.
14net use \\IP /user:… to the tool share#4Credential application.
15copy + start a.exe#1Designed copy and process-start functions.
16The AV/EDR killer runs on 40+ hosts in about two hours#7R-EXEC.
17K7RKScan.sys kills the protected security processes#2The driver serves the killer's IOCTLs: server role (R-ROLE).
18run.exe / rune.exe staged in SYSVOL, replicated by DFSR#1Designed file-write and replication functions.
19Warlock runs on 33+ hosts#7 + [DRE: Ac]Files present, but encrypted.

The full record — every step with its evidence, timestamps and rule references — is in the repository as warlock-longlegs-sharepoint-2026.json and on this site as an OKF document. It validates against the Layer-3 schema, and its notation agrees with its step sequence.

The Key Outlived the Patch

Steps 1 to 4 are where this path differs from a textbook server compromise. A one-word summary would be “SharePoint RCE”; four steps are hiding inside it.

#2 → #7 + [DRE: C] → #4 → #7
flaw   webshell reads   forged      gadget runs
       machine keys     ViewState   the commands

ASP.NET protects the ViewState with the farm's machine key: what carries a valid signature is treated as the server's own state and deserialized. The key is therefore not configuration data. It is a secret that makes the server accept input as authentic — functionally a credential. TLCTC's Axiom X separates the two events every credential has. Its acquisition is classified by the cluster that enabled it and carries the data risk event: here the webshell (#7) reading the key, [DRE: C]. Its application is always #4: the forged ViewState presented to SharePoint and accepted.

Step 4 is not a second #2. The ToolShell flaw was that SharePoint deserialized attacker input without a valid key. Once the attacker holds the key, the signed payload passes the check that exists to stop it, and deserializing it is what the platform is built to do. The weakness used in step 3 is not in the code at all: it is a secret in the wrong hands.

That has a direct consequence for remediation. A patch removes the #2 — the door the attacker used on 22 July. It does nothing to the #4: on a server that is patched but still uses the old key, the next intrusion starts at step 3. The path explains, in its own terms, the report's remark that exploitation hit “both patched and unpatched” servers, and it explains why Microsoft's ToolShell guidance in 2025 required more than the update: rotate the SharePoint ASP.NET machine keys and restart IIS on every server in the farm. Forged tokens stay valid until the key changes.

Patch closes #2. Rotation closes #4.

Whenever an attacker's code ran on a system that holds signing or decryption secrets, assume a [DRE: C] on those secrets and plan the #4 they enable. “Patched” and “clean” are different states, and only the second one ends the path.

After Day One, Almost Nothing Was Exploited

Count the clusters. Between the forged ViewState and the ransomware the attackers exploited exactly one code flaw — the driver in step 17. Everything else is either foreign code running (#7), a credential being used (#4), or a function doing what it was built to do (#1): whoami and nltest, msiexec installing from a URL, a local group taking a new member, a code editor opening a tunnel, SYSVOL replicating a folder.

Seven of the nineteen steps are #1. None of them would be stopped by a patch, because none of them uses a bug. They are stopped by restricting what functions are available, to whom and from where: whether servers may run msiexec against internet URLs, whether VS Code tunnels are allowed on servers at all, who may write to SYSVOL\scripts, and whether a change to local Administrators on a server raises an alert. That is hardening and function control — the #1 row of the control matrix, not the vulnerability-management queue.

The same pattern shows up in the Domain-Admin cascade we documented in 2025, where Storm-2603's earlier ToolShell intrusions appear as one of the entry variants: once inside, the tail of a ransomware intrusion is structurally #1, with #7 at the moments foreign code runs. The #1 cascade sets out why.

Signed Tool or Foreign Code

Two steps look alike in a SOC timeline and land in different clusters. The VS Code tunnel (step 10) is Microsoft's own, signed binary using a designed remote-development feature, and its traffic relays through Microsoft's infrastructure. Nothing foreign executes, so it is #1, and Microsoft's relay is a transit sphere — it carries the connection without being its source or its target. It is not #10: nothing becomes authoritative inside the victim's domain through a trust artifact.

#1 ||[cloud][@External⇒@Microsoft→@Org]||

NetExec (step 11) is the opposite case: an open-source offensive tool the attackers brought in. Its execution is foreign executable content, and R-EXEC requires a #7 step at that moment. What NetExec then does — LDAP queries, SMB sessions, remote commands — is #1 again. The DLL side-loading in step 7 follows the same logic: Windows' DLL search order works as designed, so side-loading is not a step of its own; it is how the attackers' code gets loaded, and the step is the code running.

The distinction matters for controls. Application control and allowlisting act on #7. They do nothing against a Microsoft-signed editor opening a tunnel — that needs a policy on the function.

The Driver Is a Server

BYOVD — bring your own vulnerable driver — looks like a local trick. Classified by role, it is a server exploit. K7RKScan.sys belongs to K7's anti-malware suite, and CVE-2025-1055 is missing access control in its IOCTL handler: any local process can ask it to terminate privileged processes, and it does. The killer (a.exe) is the client that sends the requests; the driver is the component that serves them. R-ROLE assigns the server role to the component that processes the attacker's requests, so the step is #2, with the crossing into kernel space as an intra-system annotation:

#1 → #7 → #2 |[privilege][@admin→@kernel]|
copy   killer   driver serves the
+start runs     killer's IOCTLs

Three steps, in this order. The copy and start are designed functions (#1), the killer is foreign code (#7), and the driver's flaw is exploited by it (#2). The intra-system boundary does not change the cluster (R-INTRA-7); it records where the attacker gained capability.

A correction to our own example

Our LockBit BYOVD reference path and Example 16 modelled the same technique as (#1 + #3). That was wrong three ways: a driver serving IOCTLs is in the server role (#2, the same reasoning as our CVE-2025-21333 ruling), the order is known so the steps are sequential, and the killer's execution is a #7 under R-EXEC. Both are corrected to #1 → #7 → #2, and the ransomware step there now carries [DRE: Ac] instead of [DRE: C, A].

SYSVOL: Pushing Is Not Running

The last two steps are the ones most often merged into a single word, “deployment”. The attackers copied run.exe and rune.exe into SYSVOL\[domain]\scripts\run. SYSVOL is replicated to every domain controller and readable domain-wide; on three hosts the telemetry shows the DFS Replication service (dfsrs.exe) as the process that delivered the files — ordinary replication, no separate push. Writing to SYSVOL with privileged rights and replicating it are designed functions: #1.

Execution is a different event: each host copies the files to C:\Users\Public, starts them, and Warlock encrypts. That is #7, and the data risk event is [DRE: Ac] — the files are still there, but unusable. Keeping the two apart keeps two control surfaces apart: who can write to SYSVOL and what gets replicated (a directory-hygiene control), and what is allowed to execute on an endpoint (application control).

What made the hosts start the files — a logon script or a Group Policy object — is Symantec's inference, not an observation. The path does not record it as a step.

What the Path Does Not Contain

A path is only as good as its evidence. Four things are deliberately absent:

  • Credential spraying. NetExec was used for it, but no successful logon is reported. A failed attempt is not a step.
  • Exfiltration. None is reported, so there is no [DRE: C] on business data — only on the machine keys.
  • How the domain-wide credentials were obtained. By 31 July the attackers could authenticate to almost every host and write to SYSVOL; the report does not say how they got there. Step 13 is an unresolved gap (…), with a spray success (#4) and a credential dump (#1 or #7 with [DRE: C]) as candidates. None of them can be defended on the published evidence, so none is written as a cluster.
  • The logon behind step 8. Adding an account to local Administrators on three hosts requires an administrative session on each; the step is there, marked [inferred].

The gap is not a weakness of the analysis. It is the most useful question the record raises for defenders who find the same pattern: what in your environment would answer step 13?

Velocity

The intrusion took nine days from the first webshell to encryption, and almost all of that time sits in a few transitions. Up to six days pass between the webshell and the first logged forged ViewState, about a day between the payload stage and the lateral moves, and between NetExec and the morning of 31 July. Those are VC-1 transitions (days): long enough for detection to win, if anything is looking at SharePoint's LAYOUTS folder or at msiexec fetching from catbox.

The end is a different regime. The killer reached at least 40 hosts in about two hours, and the ransomware followed through replication. Once step 14 has started, the remaining transitions are seconds to hours (VC-2 to VC-4), and replication works for the attacker at domain scale. The question for the SOC is not whether it can stop the last hour, but whether it sees the first week.

What Defenders Can Take From It

ClusterWhere it appearedWhat acts on it
#2ToolShell; the K7RKScan driverPatch SharePoint, or take it off the internet; block known-vulnerable drivers.
#7Webshell, gadget, side-loaded DLLs, NetExec, killer, WarlockAMSI integration in SharePoint, file-integrity monitoring on LAYOUTS, application control on servers.
#4Forged ViewState; admin sessions; net use with domain credentialsRotate the machine keys after any SharePoint compromise; tiered admin accounts; alert on service accounts logging on where they never do.
#1LotL discovery, msiexec from URLs, local-admin changes, VS Code tunnels, SYSVOL stagingFunction restriction and monitoring: no internet installs or dev tunnels on servers, change alerts for local Administrators and SYSVOL\scripts.

The #4 row is the one this incident adds to the usual ToolShell advice. If an organisation patched in 2025 and stopped there, the path in this report is still open to it.

Conclusion

“SharePoint exploit leads to ransomware” is accurate and not very useful. Read by cause, Warlock's path is a code flaw that was used once, a key that was stolen with it, and a week of designed functions doing what they are built to do — with foreign code running at six points and one more flaw, in a driver, at the end.

#2 → #7 + [DRE: C] → #4 → #7 → … → #1 → #7 + [DRE: Ac]
The vulnerability was patched. The credential it leaked was not.

Separating acquisition from application is not bookkeeping. It is the difference between closing the door the attacker came through and taking back the key they left with.

Further Reading


Sources. Symantec and Carbon Black Threat Hunter Team, Warlock Ransomware Attackers Hit Water and Telecom Operators, October 2026 — timeline, command lines and quotations. The Hacker News, Warlock Exploits SharePoint Flaws to Disable Security Tools and Deploy Ransomware, October 2026. Microsoft Security Response Center, Customer guidance for SharePoint vulnerability CVE-2025-53770, July 2025 — machine-key rotation. CVE-2025-1055 (K7RKScan.sys, missing access control in the IOCTL handler). TLCTC core paper (v2.6). The classification of step 13 is left open because the source does not describe the credential acquisition; step 8 is inferred from the commands that followed it, and the execution trigger after step 18 is the source's inference, as stated above.