Security teams often treat lateral movement and privilege escalation as one blur of "the attacker is spreading", but they leave different evidence, are caught by different analytics and need different containment, so an alert triaged as the wrong one is often contained in the wrong place. Lateral movement is sideways: the same or a similar level of access, used on more systems. Privilege escalation is upwards: more rights on the system, directory or cloud account the attacker already holds. This guide compares the two from a defender's point of view: how ATT&CK separates them, the telemetry each produces, how non-human identities enable both, which analytics catch which, and how containment and recovery differ. It is written for SOC analysts, detection engineers and incident responders, and deliberately stays on detection and response; hardening of remote administration channels is covered in the Lateral Movement Guide and the full technique list in the ATT&CK Privilege Escalation Tactic guide.
Key takeaways
- Sideways versus upwards. Lateral movement (TA0008) expands where an attacker can act; privilege escalation (TA0004) expands what they can do. Most intrusions alternate between them, so a confirmed instance of one is a reason to hunt for the other.
- They are seen in different places. Lateral movement shows up as relationships between hosts: logons, connections and remote execution on a destination. Escalation shows up as a change in state: a new group member, role assignment, token, service or policy.
- Containment targets differ. For lateral movement you isolate hosts and cut paths; for escalation you revoke the gained privilege and every credential that may have been exposed while it was held, because isolating one host does not undo a role assignment in the directory.
- Non-human identities enable both. Service accounts with local administrator rights on many servers are lateral movement paths; workloads allowed to pass roles, bind roles or create pods are escalation paths. Inventory them as attack paths, not just accounts.
- Valid Accounts blurs the line. T1078 sits in both the Privilege Escalation and Initial Access tactics in ATT&CK v20; a stolen privileged credential is escalation and movement at once, so identity telemetry has to be in the same investigation as host telemetry.
The short answer
Lateral movement is the set of techniques an adversary uses to enter and control remote systems. The attacker already has some credential material or code execution and uses a remote channel (SMB, RDP, WinRM, SSH, a deployment tool or a cloud control plane) to act on another system.
Privilege escalation is the set of techniques used to gain higher permissions on a system or in a directory or cloud tenant. The attacker stays on the same host or in the same account and turns user rights into administrator, SYSTEM or root, or a limited cloud role into one that can grant roles.
The two are complementary stages, not alternatives. A typical chain is: foothold as a standard user, local escalation to harvest cached credentials, lateral movement to a server where a privileged account is logged on, domain or tenant escalation, then movement to the systems that matter. Each transition is a detection opportunity, and each has a different owner in response.
Side by side
| Criterion | Lateral movement | Privilege escalation |
|---|---|---|
| ATT&CK tactic | TA0008 Lateral Movement | TA0004 Privilege Escalation |
| Direction | Sideways: more systems, similar rights | Upwards: more rights, same system or tenant |
| Typical prerequisites | Reusable credential material, a reachable remote service, rights on the destination | A local foothold plus a vulnerability, misconfiguration or over-broad permission |
| Representative techniques | Remote Services (T1021), Use Alternate Authentication Material (T1550), Lateral Tool Transfer (T1570), Software Deployment Tools (T1072) | Exploitation for Privilege Escalation (T1068), Abuse Elevation Control Mechanism (T1548), Access Token Manipulation (T1134), Account Manipulation (T1098), Domain or Tenant Policy Modification (T1484), Escape to Host (T1611) |
| Primary telemetry | Logons on the destination, east-west network flows, remote service or task creation, cloud cross-account access | Group and role changes, special privilege logons, new tokens or services running as SYSTEM, policy changes, kernel or container escape indicators |
| Best analytic shape | Sequence and graph: source host, account, destination, outcome, compared with a baseline of who administers what | State change: before and after rights, compared with an approved change or just-in-time activation |
| First containment move | Isolate source and destination hosts, block the channel, disable the account used | Remove the gained privilege, revoke sessions and tokens, reset credentials exposed while privileged |
| Scope question | Which other hosts did this account or source reach? | What did the identity do while privileged, and what can it grant? |
| Main preventive controls | Unique local credentials, segmentation, restricted admin channels, administrative tiering | Least privilege, patching, removal of escalation-capable permissions, just-in-time elevation |
Lateral movement in depth: threat model, evidence and failure modes
Threat model: the attacker holds something reusable, such as a password, an NTLM hash, a Kerberos ticket, an SSH key or a cloud access token, and the network lets the current system reach a management port or API on another system that accepts it. Pass the Hash (T1550.002), Pass the Ticket (T1550.003) and Application Access Token (T1550.001) all describe reuse of material without knowing the password.
Architecture in words: source host or workload, a credential, a channel, a destination and an outcome on the destination (a new service, a scheduled task, a remote process, an interactive session or a cloud API call). Detection is strongest when it joins all five, because any one alone is normal administration.
Evidence defenders see:
- Destination logons: Security 4624 with logon type 3 (network) or 10 (remote interactive), 4648 explicit credentials, and 4776 NTLM validation on domain controllers. Microsoft's Appendix L rates these Low criticality individually, which is why they only become useful in sequence.
- Remote execution: 4697 service installation or System 7045, 4698 scheduled task creation, and process creation under WMI or WinRM host processes, shortly after a network logon from another host.
- East-west flows: workstation to workstation connections on 445, 135, 3389, 5985 or 5986, and SSH from hosts that are not bastions. CISA and NSA list insufficient internal network monitoring among their top ten misconfigurations precisely because host-only monitoring misses movement between systems.
- Cloud and SaaS: one identity calling another account's APIs, console sessions on virtual machines, run-command style actions, or tokens used from new networks (T1021.007 Cloud Services).
Failure modes in detection: no baseline of which hosts legitimately administer which, so every admin hop looks equal; logon events collected only from servers; NTLM and Kerberos details dropped at ingestion; and cloud cross-account activity in a separate tool that nobody joins to endpoint alerts.
Privilege escalation in depth: threat model, evidence and failure modes
Threat model: three planes, each with its own evidence. On a host, an unpatched kernel or service flaw (T1068), a UAC bypass (T1548.002) or a token theft (T1134) turns a user into SYSTEM or root. In a directory, the attacker adds an account to a privileged group, edits a Group Policy Object (T1484.001) or abuses delegated rights. In cloud and Kubernetes control planes, the attacker uses a permission that lets them grant themselves more, such as adding a cloud role (T1098.003), passing a powerful role to a service or binding a cluster role.
Architecture in words: an identity with a starting set of rights, a mechanism that changes them, and a resulting state. The defensive model is a diff: what rights did this principal have before, what does it have now, and was there an approved change, ticket or just-in-time activation that explains the difference?
Evidence defenders see:
- Directory: 4728, 4732 and 4756 for members added to global, local and universal security groups. Microsoft's audit policy guidance says to monitor Administrators, Domain Admins, Enterprise Admins and Schema Admins and to alert when a standard user is added to a privileged group. Event 4964 (special groups assigned to a new logon) is rated High, meaning a single occurrence merits investigation.
- Host: 4672 special privileges at logon for accounts that should not hold them, new services or tasks running as SYSTEM created by a non-administrative parent, and EDR detections for exploitation or token manipulation.
- Entra ID: the audit activity "Add member to role (permanent)" outside Privileged Identity Management, PIM activations outside the person's usual pattern, "Update role setting in PIM", "Disable PIM alert", and the Azure Activity Log operation "Assigns the caller to user access admin", all rated High or Medium in Microsoft's security operations guide.
- AWS: role and policy attachment changes in CloudTrail. AWS notes that PassRole is not an API call, so the role passed appears only in the log of the action that created or modified the resource, such as launching an instance or creating a function.
- Kubernetes: new RoleBindings or ClusterRoleBindings, use of the escalate, bind or impersonate verbs, CertificateSigningRequest approvals and TokenRequests in the API server audit log.
Failure modes in detection: alerting on every 4672 and then suppressing the noise wholesale; no join between change tickets or PIM approvals and role changes; and blind spots where escalation happens without a role change, such as Kubernetes nodes/proxy access, which the Kubernetes project warns bypasses audit logging and admission control.
Where non-human identities enable both
Service accounts, workload identities, CI/CD runners and automation tokens are frequently the bridge between the two stages, because they hold broad standing rights and are rarely watched by a human owner.
| NHI pattern | Lateral movement it enables | Escalation it enables | Detection hook |
|---|---|---|---|
| Domain service account with local admin on many servers | One harvested password or ticket reaches every server it administers | Kerberoastable service accounts (T1558.003) in privileged groups become domain escalation | 4624 type 3 or 10 from hosts outside the account's baseline; 4769 ticket requests with weak encryption types |
| Cloud workload role with cross-account trust | Assuming roles in other accounts or projects | Roles that can attach policies, create keys or pass roles to compute | Role assumption from new sources; policy attachment by a workload principal |
| Kubernetes service account able to create pods | Reaching secrets and service accounts across the namespace | Workload creation implicitly grants the API access of any service account in that namespace | Pod creation by service accounts that never deploy; bindings created by non-CI principals |
| Deployment or RMM tool credential | Fleet-wide execution through Software Deployment Tools (T1072) | Agents usually run as SYSTEM or root on every endpoint | Bulk jobs outside change windows; new console administrators |
| Instance metadata credentials | Token reuse from outside the instance | Whatever the attached role permits | Instance credentials used from IP addresses outside the cloud provider range for that workload |
Two practical rules follow. First, model NHIs as edges in an attack graph, not rows in an inventory: the question is "where can this identity reach and what can it grant?" Second, baseline them tightly: a service account logging on interactively or from a new host is high signal. See the How to detect interactive logons by service accounts guide and the How to Find IAM Privilege Escalation Paths in AWS guide.
Detection engineering: building analytics for each
Single-event rules suit escalation; lateral movement needs sequences and baselines.
| Analytic | Stage | Logic | Main sources |
|---|---|---|---|
| Logon then execution | Lateral | Network logon from host A to host B followed within minutes by a service, task or WMI or WinRM child process on B | 4624, 4697 or 7045, 4698, process creation |
| Admin fan-out | Lateral | One account authenticates to many more hosts than its 30-day baseline in a short window | 4624, 4776, EDR network events |
| Tier violation | Both | A tier zero or Domain Admin account logs on to a workstation or member server | 4964, 4624, 4672 |
| Privileged group change | Escalation | Member added to a sensitive group with no matching change record | 4728, 4732, 4756 |
| Role grant outside JIT | Escalation | Permanent role assignment or policy attachment not issued through the approved elevation path | Entra audit logs, CloudTrail, cloud audit logs |
| Self-grant | Escalation | A principal modifies its own permissions, or creates credentials for a more privileged principal | Cloud audit logs, Kubernetes audit log |
| Audit tampering | Both (follow-on) | Audit policy change or log cleared on a host involved in either chain | 4719, 1102 |
Use Microsoft's criticality guidance as a severity anchor: High events such as 4964 and 4719 warrant investigation on a single occurrence, while Low events such as 4624 and 4672 should alert only in sequence or against a baseline. Enrich every lateral movement alert with "has this account just gained privilege?" and every escalation alert with "where has this account logged on in the last 24 hours?", so the analyst sees the chain, not half of it. Identity threat detection tooling exists largely to make that join; see the ITDR Guide.
Response: containment and eradication differ
NIST SP 800-61 Rev. 3 frames incident response as part of CSF 2.0 risk management rather than a stand-alone lifecycle, but the practical containment choices still depend on which stage you have caught. The table sets out the differences.
| Response step | Caught at lateral movement | Caught at privilege escalation |
|---|---|---|
| Immediate action | Network-isolate source and destination through EDR; block the channel between segments | Remove the new group membership or role; disable the principal or set it to deny |
| Credential action | Reset the account used and any credentials cached on both hosts | Revoke sessions and refresh tokens; rotate keys and secrets the identity could read while privileged |
| Scoping | Graph all hosts reached by the account and source in the window; look for further hops | List every action the identity took while privileged, including new accounts, keys, roles and trust changes it created |
| Persistence check | Services, tasks and tools left on destinations | Backdoor admins, federation or trust changes, new application credentials, modified policies |
| Escalation trigger | Any privileged account seen on a touched host means treat it as escalation | Domain or tenant-level privilege means assume broad lateral movement and widen scope |
| Recovery | Rebuild or clean hosts; close the path that allowed the hop | Remove the escalation path itself (patch, permission or misconfiguration); for directory compromise, plan credential resets for privileged and service accounts in a controlled order |
The most expensive mistake is the order of operations at escalation. Resetting one password while the attacker still holds a Global Administrator session or a cloud access key lets them re-grant access within minutes. Revoke the privilege and live sessions first, then rotate credentials, then hunt for persistence the privilege allowed. For cloud specifics see the Cloud Privilege Escalation Incident Playbook, and for the general trade-off between speed and evidence see Containment vs Eradication.
Worked scenario: a hybrid estate
An EDR alert fires on a finance workstation: a process spawned by the WinRM host process ran shortly after a type 3 logon from a second workstation, using a helpdesk account. That is lateral movement. The analyst enriches: the helpdesk account was added to a server administrators group 40 minutes earlier (4732) by a service account used for onboarding automation, with no change record. That is escalation, and it happened first.
- Triage: the incident is now an escalation case with movement attached. The onboarding service account is the more dangerous principal because it can change group membership.
- Contain: remove the helpdesk account from the group, disable it, isolate both workstations, and restrict the onboarding service account to deny logon while its owner confirms its normal activity.
- Scope: query every group change made by the service account in 30 days, and every host the helpdesk account reached after the group change.
- Eradicate: rotate the service account credential, convert it to a managed service account if the application allows, and remove its rights to privileged groups that onboarding never needs.
Decision criteria for detection investment
- Start with escalation analytics if: you have a large cloud or SaaS footprint, many delegated administrators, or automation that can change roles. These rules are fewer, higher fidelity and close the most damaging outcomes.
- Start with lateral movement analytics if: you run a flat on-premises network, shared local administrator passwords remain, or ransomware is the dominant concern, because spread across hosts drives impact.
- Small organisations: alert on privileged group and role changes and on Domain Admin logons to workstations, collect 4624 from all hosts, and rely on EDR for remote execution chains.
- Large organisations: add attack path analysis covering NHIs, baselines of administrative relationships, cloud and Kubernetes audit correlation, and response runbooks that switch automatically when one stage reveals the other.
Detection and measurement
- Coverage: percentage of hosts sending logon and process events, and percentage of cloud accounts, tenants and clusters with audit logs in the SIEM (target 100% for tier zero).
- Escalation fidelity: share of privileged group and role changes matched to an approved change or JIT activation; unmatched changes should trend to zero.
- Chain detection time: time from first escalation or first hop to alert in purple team exercises, tracked separately for each stage.
- Tier violations: tier zero logons to lower-tier hosts per month (target zero).
- NHI attack paths: number of non-human identities with a path to domain or tenant administrator in attack path analysis, reviewed monthly.
- Containment order: percentage of escalation incidents where privileges and sessions were revoked before credential rotation.
Framework mapping
| Framework | Lateral movement | Privilege escalation |
|---|---|---|
| MITRE ATT&CK v20 | TA0008; T1021, T1550, T1570, T1072 | TA0004; T1068, T1548, T1134, T1098, T1484, T1611, T1078 |
| NIST CSF 2.0 | DE.CM-01 network monitoring, DE.AE-03 correlation | PR.AA-05 access permissions and least privilege, DE.AE-02 analysis |
| NIST SP 800-53 Rev. 5 | SC-7 Boundary Protection, AC-4 Information Flow Enforcement, SI-4 System Monitoring | AC-6 Least Privilege, AC-6(5) Privileged Accounts, AC-6(9) Log Use of Privileged Functions, AC-2 Account Management |
| Response (both) | CSF 2.0 RS.MI-01 incidents are contained, RS.MI-02 incidents are eradicated; SP 800-53 IR-4 Incident Handling and AU-6 Audit Record Review | |
Common failure patterns
- Containing a host for a directory problem: isolating the workstation where a privileged group change was made while the added account stays in the group.
- Rotating before revoking: resetting passwords while sessions, refresh tokens or cloud keys from the privileged period remain valid.
- Single-event lateral rules: alerting on every network logon by an administrator, then suppressing the rule entirely after a week of noise.
- Ignoring the automation: treating a service account that changed group membership as trusted because "it always does that", without checking which groups it touched.
- Separate tools, separate tickets: endpoint alerts for movement and cloud alerts for role changes handled by different teams, so the chain is never seen whole.
- No log of passed roles: searching for a PassRole event in CloudTrail that does not exist instead of reviewing the resource creation call that carried the role.
Practitioner checklist
- Label every detection rule as lateral movement, escalation or both, and route it to the matching runbook.
- Collect 4624, 4672, 4697, 4698, 4728, 4732, 4756, 4964, 4719 and 1102, plus process creation, from all Windows hosts, not just servers.
- Alert on privileged group and role changes that lack a change record or JIT activation.
- Build sequence analytics joining network logon to remote execution on the destination.
- Enrich each alert with the other stage: recent privilege changes for movement alerts, recent logons for escalation alerts.
- Bring Entra, cloud and Kubernetes audit logs into the same correlation as endpoint data.
- Map non-human identities into attack path analysis and baseline their logon sources.
- Write the escalation runbook in order: revoke privilege and sessions, rotate credentials, hunt persistence.
- Test both stages in purple team exercises each quarter and record detection time for each.
Standards and references
- MITRE ATT&CK: Lateral Movement (TA0008)
- MITRE ATT&CK: Privilege Escalation (TA0004)
- NIST: SP 800-61 Rev. 3, Incident Response Recommendations and Considerations for Cybersecurity Risk Management
- CISA and NSA: Top Ten Cybersecurity Misconfigurations (AA23-278A)
- Microsoft: Appendix L, Events to Monitor
- Microsoft: Security operations for privileged accounts in Microsoft Entra ID
- AWS: Grant a user permissions to pass a role to an AWS service
- Kubernetes: Role Based Access Control Good Practices
Related NHI Mgmt Group resources: Lateral Movement Guide · ATT&CK Privilege Escalation Tactic · ATT&CK Lateral Movement Tactic · Cloud Privilege Escalation Incident Playbook · Windows Privilege Escalation Techniques Guide