MITRE ATT&CK Enterprise · TA0004
Privilege Escalation: how an attacker asks for more
Privilege EscalationTA0004
Every intrusion starts small. This is the step where a foothold becomes control — what it costs, what it looks like from the outside, and the telemetry that catches it.
For the owner
Most intrusions start small. An attacker signs in with a working password, runs a program from a shared folder, and gets away with more than they should. The moment that program asks for higher rights, the intrusion changes shape.
Nothing about that step is dramatic. It is usually a prompt on a screen, a scheduled job, a service installed to run with full authority, a group membership added to an account. From the outside, the effect is the same as a break-in that was never quiet: one person can now do things they should not, and the systems that were supposed to stop them are configured to help.
This is where small problems become expensive ones. A stolen webmail password is an afternoon. The same password attached to an account with administrative rights over the company is a company. And the escalation is often quiet on purpose, because an interruption is what brings someone running.
Three questions are worth asking your own team:
- Which of our accounts can change other accounts or run as the machine's administrator? Most organisations cannot answer this without looking, and the answer is usually longer than expected.
- What happens when a staff member or a service needs extra rights? If the answer is "they are given a shared administrator login", the pathway is open to anyone who reaches that login.
- Six months from now, can we say which account was given extra rights, by whom, and why? If we cannot, we cannot tell an escalation from routine administration after the fact.
None of this is expensive to fix. Most of it is attention, and most of the attention is spent on granting rights that were never taken back. The reason it is worth attention now is timing: the same intrusion costs far less while it is still a person with a password.
For the analyst
The tactic
MITRE ATT&CK files Privilege Escalation under TA0004. It answers a narrow question: an adversary already has a foothold — how does it acquire more authority than it was given? The matrix files 96 techniques under it, and the pinned SigmaHQ corpus carries 503 rules that name techniques in this tactic.
One property of ATT&CK matters before reading the table: a technique is filed under every tactic it serves. Valid Accounts (T1078) appears here and under Initial Access. Process Injection (T1055) appears here and under Stealth. So 503 is the number of rules that name at least one technique filed here — not the number of distinct ways an attacker can escalate, and not the number of rules you will want running.
The techniques the corpus detects
| Technique | Name | Rules in the corpus |
|---|---|---|
T1548.002 | Abuse Elevation Control Mechanism: Bypass User Account Control | 56 |
T1078 | Valid Accounts | 51 |
T1543.003 | Create or Modify System Process: Windows Service | 38 |
T1078.004 | Valid Accounts: Cloud Accounts | 37 |
T1053.005 | Scheduled Task/Job: Scheduled Task | 33 |
T1098 | Account Manipulation | 32 |
T1547.001 | Boot or Logon Autostart Execution: Registry Run Keys / Startup Folder | 31 |
T1055 | Process Injection | 26 |
T1548 | Abuse Elevation Control Mechanism | 23 |
T1068 | Exploitation for Privilege Escalation | 15 |
T1546.003 | Event Triggered Execution: Windows Management Instrumentation Event Subscription | 11 |
T1053 | Scheduled Task/Job | 9 |
T1134.001 | Access Token Manipulation: Token Impersonation/Theft | 9 |
T1543 | Create or Modify System Process | 9 |
T1053.002 | Scheduled Task/Job: At | 8 |
T1546.015 | Event Triggered Execution: Component Object Model Hijacking | 8 |
T1098.003 | Account Manipulation: Additional Cloud Roles | 7 |
T1546 | Event Triggered Execution | 7 |
T1547 | Boot or Logon Autostart Execution | 7 |
T1053.003 | Scheduled Task/Job: Cron | 6 |
T1055.001 | Process Injection: Dynamic-link Library Injection | 6 |
T1134.002 | Access Token Manipulation: Create Process with Token | 6 |
T1484.001 | Domain or Tenant Policy Modification: Group Policy Modification | 6 |
T1546.008 | Event Triggered Execution: Accessibility Features | 6 |
T1546.011 | Event Triggered Execution: Application Shimming | 6 |
T1078.002 | Valid Accounts: Domain Accounts | 5 |
T1078.003 | Valid Accounts: Local Accounts | 5 |
T1546.001 | Event Triggered Execution: Change Default File Association | 5 |
T1055.012 | Process Injection: Process Hollowing | 4 |
T1134 | Access Token Manipulation | 4 |
T1134.003 | Access Token Manipulation: Make and Impersonate Token | 4 |
T1546.002 | Event Triggered Execution: Screensaver | 4 |
T1546.007 | Event Triggered Execution: Netsh Helper DLL | 4 |
T1547.009 | Boot or Logon Autostart Execution: Shortcut Modification | 4 |
T1547.010 | Boot or Logon Autostart Execution: Port Monitors | 4 |
T1037.001 | Boot or Logon Initialization Scripts: Logon Script (Windows) | 3 |
T1078.001 | Valid Accounts: Default Accounts | 3 |
T1098.001 | Account Manipulation: Additional Cloud Credentials | 3 |
T1546.013 | Event Triggered Execution: PowerShell Profile | 3 |
T1547.004 | Boot or Logon Autostart Execution: Winlogon Helper DLL | 3 |
T1055.003 | Process Injection: Thread Execution Hijacking | 2 |
T1055.009 | Process Injection: Proc Memory | 2 |
T1543.001 | Create or Modify System Process: Launch Agent | 2 |
T1543.002 | Create or Modify System Process: Systemd Service | 2 |
T1543.004 | Create or Modify System Process: Launch Daemon | 2 |
T1546.009 | Event Triggered Execution: AppCert DLLs | 2 |
T1546.012 | Event Triggered Execution: Image File Execution Options Injection | 2 |
T1611 | Escape to Host | 2 |
T1037.005 | Boot or Logon Initialization Scripts: Startup Items | 1 |
T1055.011 | Process Injection: Extra Window Memory Injection | 1 |
T1098.005 | Account Manipulation: Device Registration | 1 |
T1098.007 | Account Manipulation: Additional Local or Domain Groups | 1 |
T1134.004 | Access Token Manipulation: Parent PID Spoofing | 1 |
T1134.005 | Access Token Manipulation: SID-History Injection | 1 |
T1484 | Domain or Tenant Policy Modification | 1 |
T1484.002 | Domain or Tenant Policy Modification: Trust Modification | 1 |
T1546.004 | Event Triggered Execution: Unix Shell Configuration Modification | 1 |
T1546.010 | Event Triggered Execution: AppInit DLLs | 1 |
T1546.014 | Event Triggered Execution: Emond | 1 |
T1547.002 | Boot or Logon Autostart Execution: Authentication Package | 1 |
T1547.003 | Boot or Logon Autostart Execution: Time Providers | 1 |
T1547.005 | Boot or Logon Autostart Execution: Security Support Provider | 1 |
T1547.006 | Boot or Logon Autostart Execution: Kernel Modules and Extensions | 1 |
T1547.008 | Boot or Logon Autostart Execution: LSASS Driver | 1 |
T1547.014 | Boot or Logon Autostart Execution: Active Setup | 1 |
T1547.015 | Boot or Logon Autostart Execution: Login Items | 1 |
T1548.001 | Abuse Elevation Control Mechanism: Setuid and Setgid | 1 |
T1548.003 | Abuse Elevation Control Mechanism: Sudo and Sudo Caching | 1 |
Five ways an intruder buys authority
Bypassing the prompt (T1548.002, T1548). Windows separates a standard process from an elevated one, and the mechanism that mediates that separation is the target. Attackers do not defeat the policy; they defeat the mechanism: a signed system binary invoked in a way that makes Windows auto-elevate it, a registry entry under the auto-elevate list, a component registered for elevation. The detection problem is that elevation is not rare. On many machines it is ordinary. What is rare is the pairing: an unfamiliar parent process spawning a binary that auto-elevates, with a command line that does not match that binary's ordinary use. The normalised columns that carry it are parent_process, process_name, command_line and user_name.
Buying authority with a configured system (T1543.003, T1543.002, T1053.005, T1053.003, T1053.002, T1547.001). A Windows service, a systemd unit, a scheduled task, a cron entry, a Run key — each is an ordinary administrative object that can run with more authority than the intruder holds. This is the most common escalation path in a small to mid-sized environment, because it requires nothing exotic: create the object, point it at your executable, let it run as the system. The telemetry is windows_sysmon process and registry events with event_type set to the relevant object, file_path showing where the service binary or task action lives, command_line showing the creation call (sc.exe, schtasks, reg add), and user_name naming the account that created it. On Linux the same idea arrives through linux_auditd execution records where user_name becomes root and parent_process is a shell.
Borrowing an identity that already has it (T1134.001, T1134.002, T1134.003, T1055, T1055.001, T1055.012). Token impersonation and theft, creating a process with another account's token, process injection by dynamic-link library injection or process hollowing. Here the attacker creates no privilege at all; they take someone else's. The telemetry is process access and remote-thread creation events, a command_line that starts a target process with someone else's token, and — the field analysts under-use — user_name on the child event. A child whose user_name differs from its parent_process owner's identity is a strong, cheap signal.
Changing the account rather than the machine (T1078, T1078.002, T1078.003, T1078.004, T1098, T1098.001, T1098.003, T1098.007, T1484.001). Account manipulation: adding a group, granting a cloud role, attaching additional credentials, modifying group policy so that rights are granted to the wrong principal. The escalation happens entirely in the identity system. The telemetry is keycloak, m365_audit, gworkspace and cloud audit records: event_type for the role or membership change, user_name of the target, src_ip of the administrative session, message carrying the change description. In many Canadian mid-sized environments this is the only place the escalation is visible at all, because the account change happens on a server that no endpoint agent watches.
Exploiting a flaw in the operating system (T1068, T1611). Kernel and driver exploits, and escaping a container or virtualised process to the host. This is the technique with the least endpoint telemetry and the greatest consequence, because the escalation is invisible to process-creation logging. Rules exist; the honest answer is that they need driver and exploit telemetry that most organisations do not collect, and they are the reason the other four paths deserve the attention.
Where the telemetry comes from
windows_sysmon— process creation, image load, registry value sets, service creation. Carriesprocess_name,parent_process,command_line,file_path,user_name.linux_auditd— execution records, privilege transitions, service and unit writes. Carriesuser_name,parent_process,command_line,file_path.keycloak,m365_audit,gworkspace,cloud— role grants, group membership changes, password resets, administrative sign-ins. Carriesevent_type,user_name,src_ip,message.firewall_syslog— rarely relevant to the escalation itself and valuable immediately after it, when the escalated identity starts talking to a new destination.
The general shape: the endpoint sees the attempt to escalate, the identity system sees the grant. Coverage of only one of the two produces a false sense of security, and the false negative below is exactly that case.
What the numbers do and do not tell you
503 rules map to this tactic because the corpus is broad and because ATT&CK files the same technique under several tactics. It does not mean 503 independent ways to be attacked.
Coverage is also uneven in a way that matters. The techniques with many rules are the ones whose artefacts are visible in ordinary telemetry — the elevation bypass, the service install, the scheduled task. The techniques with one or two rules are the obscure corners, and those are where a genuinely novel intrusion would live. What the corpus gives you is a floor, not a ceiling.
Detection engineering notes
- Anchor on the identity change, not on the tool. "A scheduled task was created" fires on every deployment and every patch cycle. "A scheduled task was created by an account that has never created one" fires rarely and means something.
- Parent-child relationships are the cheapest high-value signal. Elevation almost always arrives through a parent that does not normally spawn it: a document handler, a browser, a scripting interpreter, a service manager.
- Watch the user, not only the binary. A privileged
user_nameappearing on a host it has never touched, or a standard account becoming privileged mid-session, is rarer and stronger than any list of executable names. - Baseline your own administration or drown in noise. Software deployment, patch agents and backup products create services, tasks and Run keys legitimately. A rule set that fires on all of them is muted, and a muted rule set protects nothing.
- Correlate forward, not only backward. The strongest signal is often the escalation followed by behaviour the escalated identity has never shown: a new source address, a new destination, a new file location.
Where analysts get this wrong, and what a false negative looks like
The most common error is treating a rule count as coverage. A rule that needs an event stream you do not collect is not a detection; it is a document. The second common error is assuming escalation means full control of a machine. In a mid-sized organization the escalation that matters is often a cloud role, a password reset or membership of a distribution group, and none of those touch an endpoint.
A false negative looks like this: an intruder who already holds a valid administrator session requests one more right through the ordinary help-desk process. Every tool fired exactly as designed. The request came from an address inside the corporate range, from an account with a legitimate history of administering, on a Tuesday afternoon. The escalation is visible in the identity audit trail and nowhere else, and it looks like Tuesday.
Where DetectMonkey fits
Privilege escalation is the step where a chain of small events becomes one expensive event, which makes it a detection-latency problem as much as a coverage problem. The escalation attempt, the object created to carry it, and the identity change that authorised it can all arrive within seconds of each other — but only if the endpoint stream and the identity stream are evaluated together, as the events arrive rather than after they are indexed. A detection that lands an hour later is a different product than one that lands in the first minute, because the hour is when the intruder takes the actions you will have to explain.
DetectMonkey is built for that shape of work. The collectors carry endpoint process, registry and service events alongside identity audit records, the pinned corpus of 3164 detection rules is evaluated as the events arrive, and every alert arrives with the evidence that produced it, in English and French. The platform detects and alerts. It does not isolate, quarantine, block or remediate anything on your systems, and it does not change your configuration. Telemetry is processed and stored on private servers operated by Codein in New Brunswick, Canada — not in a public cloud.
Early access is 90 days at no charge, for up to 100 seats, with founder-led onboarding and no card required.
See it on your own telemetry
Early access is 90 days at no charge, for up to 100 seats, founder-led onboarding, no card required.