Revised for TLCTC v2.7 on 25 October 2026 (from the April 2026 original, which was written on v2.1). I re-checked the incident facts against the primary sources. The Lynx case is the DFIR Report of December 2025, and its exfiltration ran through a browser upload, not attacker tooling. Microsoft's Akira case names no threat actor, so the earlier “Storm-0300” label was wrong, and that encryption was blocked. Talos never observed Storm-2603's initial access. The doctrine is also updated. Every cluster step records its own SRE. Accounts the attacker created are not #4 when used. Attacker-introduced tools execute as #7. Velocity classes are read per transition, ZeroLogon's DRE has moved off the exploit step, and the right side of the bow-tie follows the v2.7 consequence chain.
Practitioners often describe Active Directory ransomware attacks as a mass of overlapping techniques: “lateral movement,” “defense evasion,” “impact.” These labels name what the attacker is trying to accomplish. They do not say which generic vulnerability the attacker exploits at each step, so they do not say which control would have stopped it. A TLCTC decomposition does. Run it against real 2025 incidents and a pattern appears. Once the attacker holds Domain Admin, almost every remaining step is #1 Abuse of Functions.
This is no accident. It follows from what a Tier-0 role is. The cascade below shows why, using three 2025 cases as ground truth: Lynx (DFIR Report), an attempted Akira attack (Microsoft) and Storm-2603 (Cisco Talos).
1. The invariant pattern
Microsoft reported in April 2025 that threat actors breach a domain controller in more than 78% of the human-operated cyberattacks it sees. In more than 35% of cases the domain controller is the primary “spreader device” that distributes ransomware at scale. Verizon's 2025 DBIR found ransomware in 44% of breaches, up from 32% (the count includes non-encrypting extortion), and in 88% of breaches at small and medium-sized businesses. Speed is the other constant. Secureworks' 2023 State of the Threat report put the median time from initial access to ransomware deployment under one day, down from 4.5 days a year earlier, and within five hours in 10% of engagements.
Tooling, ransomware family and initial vector all vary. The bow-tie shape stays the same:
prefix (variable) → decisive compromise: a Tier-0 identity under attacker control → tail (near-invariant):
#1 … #1 + [DRE: C] … #1 + [DRE: Av] → #1 → #7 + [DRE: Ac] (per host)
“Decisive” is deliberate, and it is not “the only”. In TLCTC v2.7 every successful cluster step records its own System Risk Event: a path of n cluster steps records n System Compromises (core §3.4). The Domain Admin takeover marks the point where the attacker's reach becomes domain-wide. Loss of control began earlier, at the first successful step.
The prefix is where clusters genuinely diverge: initial access, foothold, escalation. The tail is the part most defenders call “the ransomware attack,” and it is nearly identical across families. It is almost entirely #1. #7 appears only where foreign executable content runs: attacker tools, persistence agents, the exfiltration script in some cases, and the encryption payload on each victim host.
2. The prefix: five routes to Domain Admin
The route from outside to Domain Admin in hand is the variable part. Three routes come from the 2025 cases. The other two are recurring patterns with well-known forensic signatures.
Variant A: valid-credential RDP (Lynx, March 2025)
The DFIR Report's Lynx case (published December 2025) is instructive for what is absent. There was no brute-force noise and no failed authentication, just a single successful RDP logon to an internet-exposed host with valid credentials. The actor dropped SoftPerfect NetScan through the RDP session and ran it. About ten minutes after the first logon, they moved over RDP to a domain controller with a different domain admin account. The report found no on-host privilege escalation and no credential dumping. Both credentials were most likely compromised before the intrusion began, through infostealers, reuse from a breach or an initial-access broker.
… → #4 →[Δt=mins] #7 (NetScan) →[Δt=mins] #4 (domain admin to DC, ~10 min after first logon)
The acquisition prefix is an unresolved gap (…) because no acquisition telemetry survived on the target estate (R-UNRES-8). Each logon with a stolen credential is #4 (R-CRED). NetScan is a legitimate product, but here it is a foreign-introduced tool running at the attacker's command, so its execution is #7 (R-EXEC).
Variant B: VPN foothold, then a domain admin credential (Akira, Microsoft case study)
Microsoft's April 2025 case study describes an attempted Akira attack on a small-to-medium manufacturer. Microsoft names no threat actor. The attacker came in “presumably” through the VPN. They mapped assets and tried to escalate, and Defender caught a wide remote secrets dump and contained the first account. The attacker then obtained a second user's domain admin credentials, possibly from devices Defender did not cover, and connected over RDP from an attacker-controlled device to DC1. On the DC they mapped servers and shares, changed security-related Group Policy settings to tamper with antivirus, and created two new accounts, adding both to Domain Admins.
#4 [conf=low] (VPN) → … → #4 (RDP to DC1, domain admin) → #1 (recon) → #1 (GPO: AV tampering) → #1 (create two accounts, add to Domain Admins) → #1 (log on as one of them)
The VPN entry is #4 [conf=low] because a cluster can be defended, if only weakly (R-UNRES-9). The joint Akira advisory (CISA AA24-109A, updated November 2025) names both VPNs without MFA and exploited VPN flaws, such as Cisco CVE-2020-3259 and CVE-2023-20269 and SonicWall CVE-2024-40766. If an appliance flaw was the way in, the first step is #2. The path then ends, because Microsoft's automatic attack disruption contained the accounts and the RDP device and blocked the remote encryption from DC1. Only successful steps are recorded, so there is no #7 + [DRE: Ac] here. The last step is #1, not #4, because the account was one the attacker had created (see the R-CRED reminder below).
Variant C: server exploit, then Tier-0 accounts (Storm-2603, August 2025)
Cisco Talos responded to a Storm-2603 intrusion in August 2025 (attribution: moderate confidence) and published it in October 2025. Talos could not observe initial access. It judged ToolShell exploitation of on-premises SharePoint likely, given the victim's exposure and the attribution. Microsoft had already documented Storm-2603 exploiting ToolShell to deploy Warlock in July 2025. Inside, the actor created admin accounts that synced to Entra ID and used one of them to reach the VMware vSphere console. They changed Active Directory Group Policy to turn off Defender's real-time protection and behavior monitoring, exfiltrated documents with a PowerShell script (HTTP PUT to an external server) and encrypted with a fileless PowerShell encryptor. LockBit-identified binaries and Warlock-style .xlockxlock files appeared on Windows, and Babuk partially encrypted ESXi.
#2 [conf=low] (ToolShell) → … → #1 (create admin accounts, synced to Entra ID) → #1 (vSphere console) → #1 (GPO: Defender off) → #7 + [DRE: C] (PowerShell exfil script) → #7 + [DRE: Ac] (encryption)
The server-side flaw is #2 per R-ROLE, carried at low confidence because it was not observed. The Entra ID sync is the organization's own designed pipeline, so it adds reach but no step. The actor also installed Velociraptor 0.73.4.0, a legitimate DFIR tool, on several servers for persistence. Each install is a #7 (foreign-introduced tool). Talos gives no timing for them, so they stay out of the ordered line. That version is vulnerable to CVE-2025-6264, but Talos could not determine whether the flaw was exploited, so no #2 is recorded for it. The ToolShell entry chain itself is decomposed in The Key Outlived the Patch.
Variant D: Kerberoasting (pattern)
Requesting service tickets for accounts with a service principal name is a designed Kerberos function, and any authenticated principal may do it. The ticket is encrypted with a key derived from the service account's password. The attacker takes it offline and cracks it. A weak password falls in minutes to hours, and a long managed password does not fall at all. The recovered password is then used.
#4 (foothold account) → #1 + [DRE: C] (TGS requests for SPNs) →[Δt=~hours] #4 (cracked service account) → #1
Forensic signature: a burst of Event ID 4769 with TicketEncryptionType = 0x17 (RC4-HMAC) for user-account SPNs from one source. Offline cracking happens on no system of the target, so it is no step. The Δt across it is the cracking time. Per Axiom X and R-CRED, the [DRE: C] sits on the acquisition step (#1) and never on the #4 use. This follows the ATT&CK mapping for T1558.003 (#1 → #4).
Variant E: ZeroLogon (CVE-2020-1472)
Netlogon (MS-NRPC) computed its client credential with AES-CFB8 under a fixed all-zero IV. For about one session key in 256, an all-zero challenge yields an all-zero credential, so an unauthenticated attacker on the network authenticates as any computer after a few hundred tries. Secura's write-up puts that at about three seconds. The attacker then calls the designed password-change function to set the DC's machine-account password to empty, logs on as the DC and runs DCSync. Microsoft patched it in August 2020 and enforced secure RPC in February 2021. CISA's KEV catalog lists known ransomware campaign use.
#2 + [DRE: I] (DC machine-account password reset) →[Δt=s] #4 (log on as the DC account) →[Δt=s] #1 + [DRE: C] (DCSync: every hash)
The flaw sits in the domain controller acting in server role, hence #2 per R-ROLE. The exploit step discloses nothing. It overwrites the machine-account password in the directory, a record changed outside its authorized process ([DRE: I]). The hashes are disclosed by DCSync, a #1 step. The April edition of this post put a [DRE: C] on the exploit step, which credited one step with what a later step did. “Privilege escalation” is the effect, not a cluster (Axiom III).
3. The post-DA tail: the #1-cascade in detail
This is where the structural insight lives. Almost every step below uses a designed, documented capability of Windows, Active Directory or the backup software. No implementation flaw is needed, and no specific generic vulnerability (capacity, channel, substrate) is the one exploited. That is the residual test for #1 under R-SPECIFIC. The rows that are not #1 are the ones where foreign executable content runs. The Seen in column marks which of the three cases evidences the row: L = Lynx, A = Akira (Microsoft), S = Storm-2603 (Talos). A dash marks a recurring pattern that none of the three documents.
| # | Operator action | Mechanism | TLCTC | Seen in | Evidence |
|---|---|---|---|---|---|
| 1 | Discovery with native tools | net, nltest, systeminfo, dsa.msc |
#1 | L, A | 4688; 4662; LDAP query logs |
| 2 | Discovery with attacker-introduced tools | NetScan, NetExec, SharpHound | #7 | L | 4688; Sysmon 1/3; browser download history |
| 3 | Create look-alike or admin accounts | dsa.msc, net user /add, New-ADUser |
#1 | L, A, S | 4720; 4738 |
| 4 | Add to Domain Admins, Group Policy Creator Owners | dsa.msc, net group, Add-ADGroupMember |
#1 | L, A | 4728; 4732; 4756 |
| 5 | Persistence agent installed | AnyDesk service, Velociraptor | #7 | L, S | 7045; 4688 |
| 6 | DCSync (NTDS replication pull) | MS-DRSR GetNCChanges |
#1 + [DRE: C] | – | 4662 with DS-Replication-Get-Changes-All GUID |
| 7a | LSASS dump with built-in means | comsvcs.dll MiniDump via rundll32, Task Manager |
#1 + [DRE: C] | – | Sysmon EID 10; 4688 |
| 7b | LSASS dump with a brought-in tool | Mimikatz, ProcDump | #7 + [DRE: C] | – | 4688; AMSI; Sysmon EID 10 |
| 8 | Create or edit a GPO | New-GPO, GPMC |
#1 | A, S | 5136; 5137; SYSVOL writes |
| 9 | Disable Defender via GPO | “Turn off real-time protection”, behavior monitoring off | #1 | A, S | 5136; Defender 5001/5007 |
| 10 | Collect and archive files | SMB share reads, 7-Zip “Add to archive” | #1 | L | 5140; 5145; 4688 (7zG.exe) |
| 11a | Exfiltration through a designed client | Browser upload to temp.sh | #1 + [DRE: C] | L | Browser history (/upload); proxy logs |
| 11b | Exfiltration by an attacker script | PowerShell HTTP PUT loop | #7 + [DRE: C] | S | 4104; proxy/egress logs |
| 12 | Backup job deletion | Veeam Backup & Replication console | #1 + [DRE: Av] | L | Veeam job-deletion log; 4624/4672 on backup server |
| 13 | Shadow copy deletion | vssadmin delete shadows /all, wmic shadowcopy delete |
#1 + [DRE: Av] | – | 4688 (vssadmin); VSS 8224 |
| 14 | Payload staging | Copy to a desktop or ADMIN$/C$, GPO file deployment |
#1 | L | 5140; 5145; file creation |
| 15 | Remote execution trigger | Scheduled task (GPO), service creation, wmic process call create |
#1 | – | 4698; 4702; 7045 |
| 16 | Encryption payload runs (per host) | PE binary (w.exe), fileless PowerShell, ESXi ELF |
#7 + [DRE: Ac] | L, S | 4688; 4104; rename storm; ransom note |
| 17 | Log clearing (anti-forensics) | wevtutil cl, Clear-EventLog |
#1 + [DRE: Av] | – | 1102; 104 |
In the Akira case the GPO tampering (A) succeeded but the encryption (row 16) was blocked, so the row does not apply there. Microsoft's “wide remote secrets dump” in that case was detected and contained. Whether it succeeded is not reported, so it is not counted against row 7.
R-EXEC reminder. Every execution of foreign executable content is its own #7 step. This includes legitimate products the attacker brings in (NetScan, NetExec, AnyDesk, Velociraptor), because #7 covers “foreign-introduced tools”. Each target host gets its own #1 → #7 pair. A forensic record writes each pair out with its own [DRE: Ac]. A compact “×N” fan-out marker exists only as a notation proposal, not as canon.
R-CRED reminder. Rows 6 and 7 are acquisition, and the [DRE: C] attaches there. Later use of a stolen credential to authenticate (pass-the-hash, pass-the-ticket, a forged Golden Ticket) is always #4, and it records no DRE of its own. The opposite case matters just as much. An account the attacker created (row 3) was issued to the attacker by the directory's own enrolment function, so the attacker is its authentic holder. Logging on with it is not #4. The exploited weakness was the enrolment, already recorded as #1. Lynx's look-alike logons to the hypervisors on day 2 and Akira's second logon to DC1 are both #1.
Storm-2603 shows the end of the tail in its purest form. Group Policy switched off Defender's real-time protection, then a fileless PowerShell script did the encrypting. “Fileless” describes disk persistence. It does not change the classification. powershell.exe is a designed host process. The attacker's script running inside it is foreign executable content (#7), and the execution is recorded as its own step whether or not anything was written to disk. The same holds for the exfiltration script in the same incident. Compare Lynx, where the data left through a browser's upload form, a designed client function. That is #1, and there is no #7 to record.
4. Reference path: Lynx, evidence only
The full Lynx intrusion in TLCTC notation, limited to the steps the DFIR Report evidences:
… → #4 (day 1: RDP to the beachhead, valid credentials) →[Δt=mins] #7 (NetScan dropped over RDP and run) →[Δt=mins] #4 (RDP to a DC as a second domain admin, ~10 min after first logon) →[Δt=mins] #1 (create "administratr" + a look-alike with dsa.msc) →[Δt=mins] #1 (add to Domain Admins, Group Policy Creator Owners) →[Δt=mins] #7 (AnyDesk installed as a service, never used) →[Δt=~5d] #4 (day 6: RDP back to the beachhead) →[Δt=mins] #1 → #7 (NetExec downloaded via Edge, run over SMB) →[Δt=mins] #1 (7-Zip archives from two file servers) →[Δt=mins] #1 + [DRE: C] (browser upload to temp.sh) →[Δt=~3d] #4 (day 9: RDP to the beachhead from a new IP) →[Δt=mins] #1 + [DRE: Av] (Veeam console: backup jobs deleted) →[Δt=mins] #1 → #7 + [DRE: Ac] (w.exe copied and run via cmd, per backup and file server)
Compressions: the day-2 look-alike logons to hypervisors (#1, per the R-CRED reminder) and the day-8 reconnaissance of domain controllers and Hyper-V hosts are left out. Only the session-opening logons are shown, although every RDP hop with the stolen domain admin credential is its own #4. The report does not mention shadow-copy deletion or log clearing, so neither appears.
Time to ransomware: just under 178 hours across nine calendar days. Velocity is a property of each transition, not of the incident. Inside each session the transitions are VC-3 (minutes). Between sessions the gaps are VC-1 (days). The edge that matters most is the ten minutes from the first logon to a domain controller. A transition at VC-3 is too fast for purely human response (core §7.2). Containment at that edge needs DCS_c = TTC_P90 / Δt below 1, that is, a P90 time-to-contain under ten minutes. Microsoft's April 2025 post claims automatic containment of a compromised domain controller in under three minutes. That kind of automated response is what this transition requires. Faster operators shrink the Δt values. The step sequence stays the same.
The gap notation lets analysts record the structure of what happened without inventing steps. Use … for an unknown prefix, and annotate it with what is unresolved (R-UNRES-8). Use ? for a single unresolved step, and #1 → … → #1 where the attacker's log clearing destroyed coverage. Wherever any cluster can be defended, write #X [conf=low] instead (R-UNRES-9), as for Akira's VPN entry and Storm-2603's ToolShell. Paths are living artifacts.
5. Why the post-DA phase is structurally #1
Every tool the attacker picks up after taking Domain Admin is software working within its designed scope. That covers GPMC, dsa.msc, vssadmin, net, scheduled tasks, WMI, ADMIN$, directory replication and the backup console. None of it needs an implementation flaw. What the attacker exploits is exactly the generic vulnerability of #1: “The inherent trust, scope, and complexity designed into software functionality and configuration.” A Tier-0 role holds the largest share of that designed scope in the estate. That is why the tail is #1 almost end to end (Axiom VII: each vector is defined by the generic vulnerability it initially targets).
Why is it an attack at all, when the functions are designed to be used? R-SCOPE answers this. A step is classified under a cluster when the action fell outside the entitlement an accountable grantor actually conferred for it. An entitlement attaches to the person, not to the token. Using a credential by anyone other than its grantee is never inside an envelope. That makes it #4, and the function use that follows is #1. The attacker holding the domain admin's credentials holds none of the domain admin's entitlements.
Same event IDs, different row. Suppose the backup administrator deletes the Veeam jobs on purpose, inside the grant the organization gave them. That is Abuse of Rights: operational risk, no cluster and no System Risk Event, and the consequence chain starts directly at the [DRE: Av]. The log entry looks identical to row 12 of the table. What differs is whether an accountable grantor conferred the entitlement for this action. Two different control regimes answer the two rows: identity and function controls for the attack, governance of the grant itself for the abuse.
That is why you cannot “prevent ransomware” by mapping one control to the encryption outcome. The controls have to map to the steps. For the #1 steps in the tail, that means GPO change auditing, VSS and backup tamper protection, Protected Users, Tier-0 isolation, backup-console MFA with out-of-band approval, ITDR on replication anomalies, canary SPNs and privileged access workstations. For the #7 steps, it means EDR behavioral detection, ASR rules and application allowlisting, including for “legitimate” remote-access and DFIR tools nobody approved. For the #4 steps, it means detecting stolen credentials and requiring MFA at every edge. Labels like “lateral movement” or “defense evasion” merge four to six distinct #1 steps, each with its own control surface. That merging is why “anti-ransomware” as a product category underperforms.
6. The right side of the bow-tie
The cascade ends with three data risk events of different kinds. The [DRE: C] is disclosure: business files on temp.sh, or credential material from DCSync. The [DRE: Av] is data that is gone: backup jobs, shadow copies, logs. The [DRE: Ac] is data still present but unusable: the encrypted file servers. TLCTC v2.7 states how they reach the business: SRE → DRE → BRE*, through two kinds of edges (core §3.4; Layered Event Modelling §6).
| From | Edge | To (Business Risk Event) | Δt? |
|---|---|---|---|
| [DRE: Ac] file and backup servers | causal | File services and dependent processes stop; outage declared | yes: the causal interval, which bounds the window for detection and response |
| [DRE: Av] backup jobs | causal | Restore from backup fails; recovery falls back to rebuild | yes |
| [DRE: C] personal data among the files | normative | Notification duty arises (e.g. GDPR Art. 33) | no: the 72 hours run from awareness and bound a separate response process |
| Outage declared | causal | Further BREs (lost revenue, contract penalties) up to the terminal BRE that risk appetite names as Business Impact | yes |
Two consequences for defenders follow. First, the chain can break only at a causal edge. The backup deletion and the encryption are separate causal routes to the outage, which is why backup immutability is a control on the right side of the bow-tie and not merely hygiene. Second, the notification duty carries no Δt. It arises as soon as personal data is disclosed, however fast or slowly anyone notices. The 72-hour deadline measures the organization's response from awareness, not the attacker. The full ladder is in The Cyber Bow-Tie.
7. Practical notation tips
- Domain Admin takeover is the decisive SRE, not the only one. Every successful cluster step records a System Compromise. The Tier-0 takeover is where reach becomes domain-wide. Mark it in diagrams as an event, never as a cluster.
-
Do not classify “privilege escalation” as a cluster. It is an effect (Axiom III). When Domain Admin comes from ZeroLogon, the path is #2 → #4 → #1. When it comes from Kerberoasting, it is #1 → #4. An intra-system marker such as
|[privilege][@user→@system]|may annotate a local escalation, but it never changes the cluster (R-INTRA-7). - Accounts the attacker created are not #4. Creation is #1 (enrolment outside any grant). Later logons with those accounts are #1 too, because nobody else's identity is claimed (R-CRED). Only stolen or forged credentials make #4.
- Backup destruction is dual-DRE territory. Deleting backup jobs or shadow copies is [DRE: Av]: the data is gone. Encrypting primary data is [DRE: Ac]: the data is present but unusable. Keep them separate. They have different recovery semantics and call for different controls.
- Exfiltration is #1 or #7, decided by what runs. A browser or other designed client uploading the files is #1 + [DRE: C] (Lynx). An attacker script or binary doing the upload is #7 + [DRE: C] (Storm-2603).
-
Anti-forensics is not out of scope.
wevtutil clis #1 + [DRE: Av] on log data. Where it leaves gaps, the…notation earns its keep: #9 ||[human][@External→@Org]|| →[Δt=18h] #4 →[Δt=2m] #1 → … → #1. The boundary operator is required on a bridge cluster such as #9. - Blocked steps are not steps. A cluster assignment asserts that the step succeeded. Microsoft's Akira case ends before encryption because the encryption was blocked. Record the attempt in prose, never as #7 + [DRE: Ac].
- “Fileless” does not mean uncategorized. A PowerShell script running in memory from a GPO-scheduled task is still foreign executable content. The task trigger is #1, and the attacker's script is #7. Record both.
Once you see the cascade, you can't unsee it. Calling post-DA activity a diffuse cloud of “TTPs” hides a crisp structural fact. Holding a Tier-0 credential turns the rest of the attack into a sequence of designed function calls, each one working exactly as built and none of them entitled. The defender's job is to intercept the cascade before Domain Admin, or to detect and contain it within the minutes the transitions allow after it. Once it is running, every step looks like administration.
References
- The DFIR Report, Cat's Got Your Files: Lynx Ransomware, 17 December 2025 (intrusion of March 2025).
- Microsoft Threat Intelligence, How cyberattackers exploit domain controllers using ransomware, 9 April 2025.
- Microsoft Threat Intelligence, Disrupting active exploitation of on-premises SharePoint vulnerabilities, 22 July 2025.
- Cisco Talos, Velociraptor leveraged in ransomware attacks, 9 October 2025 (incident of August 2025).
- CISA et al., #StopRansomware: Akira Ransomware (AA24-109A), April 2024, updated 13 November 2025.
- Secura (T. Tervoort), Zerologon: Unauthenticated domain controller compromise by subverting Netlogon cryptography (CVE-2020-1472), September 2020; CISA Known Exploited Vulnerabilities catalog, CVE-2020-1472.
- Verizon, 2025 Data Breach Investigations Report; Secureworks, State of the Threat Report 2023.
- TLCTC v2.7 core: §3.4 (SRE and consequence chain), §3.5 (cause-side partition, R-SCOPE), §7.2 (velocity, DCS), §7.6 (DRE tags), §7.7 (epistemic states). Glossary: SRE, BRE, R-SCOPE, Abuse of Rights, R-CRED, velocity class, DCS.
TLCTC Framework · v2.7 · CC BY 4.0