Lateral movement is the stage of an attack in which an intruder who has gained a first foothold uses it to reach other systems, accounts and data. The first foothold is rarely the goal. A stolen key or a compromised server is a starting point, and the attacker’s next job is to move sideways, and then up, until they reach something worth taking.
MITRE ATT&CK treats it as its own tactic, Lateral Movement (TA0008), and lists techniques such as using valid accounts, reusing stolen tokens and exploiting remote services. In traditional networks, this meant hopping between servers. In the cloud it often means something different and faster: using the permissions attached to one identity to act on other resources, without touching a network at all.
This page explains how lateral movement works in cloud environments, the common patterns, how it shows up in logs, and the controls that limit it. It connects directly to blast radius and attack path analysis.
Why cloud lateral movement is different
On a classic network, an attacker moves by connecting from machine to machine, stealing passwords from memory and exploiting services. Those steps still happen, for example when an attacker pivots between virtual machines or Kubernetes pods. But the cloud adds a second layer of movement through the control plane, the APIs that create, change and read cloud resources.
- Identities are the connections. A role that is allowed to read a storage bucket, assume another role or fetch a secret is a path. No network route is needed.
- Everything is reachable by API. If an identity has permission, it can act from anywhere that can reach the cloud provider’s API.
- Permissions chain. One role may be allowed to assume another, which can read a secret, which holds a database password. Each hop looks like normal activity.
- Machines hold credentials. Compute instances, containers and functions are given temporary credentials by the platform, so compromising a workload gives the attacker an identity.
This is why a single leaked key can produce a large incident. The attacker follows permissions, and the size of the breach is set by how far those permissions reach.
Common cloud lateral movement patterns
1. Credential chaining
The attacker finds a credential inside the first system, uses it to enter a second, finds another, and repeats. Barracuda reported this pattern in November 2026: after an attacker used a leaked vendor key, they used credentials found in one system to move through data analytics, source-code repositories and Kubernetes clusters. See Barracuda’s cloud breach.
2. Role assumption and trust abuse
Cloud roles can be assumed by other identities if a trust policy allows it. An attacker with permission to assume a role gains that role’s access. Overly broad trust, such as a role that trusts an entire account or a wildcard, makes this easy.
3. Metadata service credential theft
A compromised workload can ask the instance metadata service for the temporary credentials attached to its role. If the role is broad, the attacker inherits it. See the cloud instance metadata service (IMDS).
4. Secrets and configuration stores
Secrets managers, parameter stores, environment variables and configuration files hold database passwords and API keys. An identity allowed to read them can pivot to the systems they unlock.
5. Privilege escalation along the way
The attacker looks for permissions that let them grant themselves more, such as creating access keys, attaching policies or modifying a role. See privilege escalation detection.
6. Container and Kubernetes pivots
From a compromised pod, an attacker may reach the node, steal service account tokens, or call the cluster’s API. Mounting the host’s file system into a container, which Barracuda observed in an attempt to persist, exposes credentials on the host.
7. Cross-account and cross-tenant movement
Organizations with many accounts use cross-account roles. A compromise in a low-trust account can reach a high-trust one if the trust is too wide. Managed service providers and software vendors with access into customer tenants create a similar route.
8. CI/CD and developer tooling
Build pipelines often hold powerful credentials. Compromising a repository, a runner or a developer’s editor extension can yield access to production. See Cursor’s breach via a vendor key.
9. Application and token reuse
Stolen session tokens and application access tokens allow movement without passwords. MITRE lists these under Use Alternate Authentication Material (T1550). See session hijacking.
Why speed matters now
Lateral movement used to be slow because a human had to look around, think and try things. AI-assisted attack tools remove much of that delay. The FBI warned in November 2026 that one such framework, ARTEX, has taken some intrusions from first targeting to privileged access in under three hours. See the FBI warning on ARTEX and AI-powered attacks and autonomous penetration testing. When movement is automated, the time between the first foothold and a serious breach shrinks, and defenses that rely on a person noticing quickly lose ground.
How to detect it
Because the attacker uses valid credentials and legitimate APIs, detection depends on behavior and context rather than known-bad signatures.
| Signal | What it suggests |
|---|---|
| An identity calling APIs it has never used | Exploring what it can do |
| Bursts of access-denied errors | Probing for permissions. See below. |
| A role assumed from an unusual source or by an unusual identity | Role-chaining |
| Reading many secrets or parameters in a short time | Credential harvesting |
| New access keys, users or policy attachments | Persistence and escalation |
| A workload using its credentials from outside the cloud | Stolen metadata credentials |
| Activity from a new geography or at unusual hours | Compromised identity |
| Sudden enumeration calls (list, describe, get) across many services | Reconnaissance |
Example for AWS CloudTrail logs in CloudWatch Logs, finding identities with unusual numbers of denied calls:
fields @timestamp, userIdentity.arn, sourceIPAddress, eventName, errorCode
| filter errorCode in ["AccessDenied", "UnauthorizedOperation", "Client.UnauthorizedOperation"]
| stats count() as denied by userIdentity.arn, sourceIPAddress, bin(5m)
| sort denied desc
| limit 50
And role assumptions, to spot unexpected chains:
fields @timestamp, userIdentity.arn, requestParameters.roleArn, sourceIPAddress
| filter eventName = "AssumeRole"
| stats count() as assumptions by userIdentity.arn, requestParameters.roleArn
| sort assumptions desc
See also the lateral movement detection use case, cloud UEBA and cloud detection and response.
How to limit it
The aim is to make each hop harder, shorter and more visible.
Shrink what any one identity can reach
- Apply least privilege to roles, users, service accounts and agents. Remove unused permissions.
- Tighten trust policies. Trust specific roles and accounts, not wildcards. Use conditions such as external IDs, source addresses and required MFA where they apply.
- Use permission boundaries and organization policies as ceilings that cannot be exceeded. See AWS IAM permission boundaries.
Shorten how long access lasts
- Remove standing privilege. Access that exists only inside an approved window gives a chaining attacker less to reuse. See zero standing privilege and just-in-time access.
- Use short-lived credentials for machines and pipelines instead of long-lived keys. See orphaned credentials.
Reduce the places credentials hide
- Keep secrets out of code and configuration. See secrets sprawl.
- Require the newer, session-based version of the instance metadata service and restrict workload roles.
- Separate environments and accounts, so a compromise in development cannot reach production.
Add friction and visibility
- Segment networks and clusters, and restrict east-west traffic.
- Log control-plane and data access and keep logs off the systems being watched. See AWS CloudTrail.
- Alert on the signals above and automate the safe responses, such as disabling a credential.
Know your paths
You cannot defend paths you have not mapped. Attack path analysis shows how an identity or workload connects to sensitive data, so you can cut the highest-value routes first.
How Cloudanix helps
Cloudanix maps the paths, removes standing access along them, and watches for movement.
- Mapping the paths. The attack path view and unified asset graph connect identities, workloads, roles and data, so you can see how a foothold reaches sensitive systems. CIEM shows which identities hold powerful or excessive permissions. CSPM ranks the findings by exposure, environment, data sensitivity and identity with contextual severity. See correlating identity with misconfigurations.
- Removing what an attacker would chain. Cloud JIT replaces standing access to AWS, Azure and GCP with time-boxed grants that revoke themselves. Database JIT, VM JIT and Kubernetes JIT do the same for databases, servers and clusters, and Agentic JIT and Coding Agent JIT give pipelines and AI agents credentials that last for a task.
- Noticing movement. Cloud UEBA and cloud detection and response look for identities acting unlike themselves.
The boundaries: Cloudanix does not block network traffic between your workloads, and it does not stop an attacker’s first foothold. It narrows the permissions and the time an attacker can use, and shows the paths, so each hop gets harder.
Further reading
- MITRE ATT&CK: Lateral Movement (TA0008)
- MITRE ATT&CK: Use Alternate Authentication Material (T1550)
- MITRE ATT&CK: Valid Accounts, Cloud Accounts (T1078.004)
Frequently asked questions
What is lateral movement?
It is the phase of an attack in which an intruder uses an initial foothold to reach other systems, accounts and data inside the environment.
How is lateral movement different in the cloud?
In the cloud it often happens through permissions and APIs, such as assuming roles and reading secrets, instead of connecting between servers. No network path is needed, only a permission.
How do attackers move laterally in AWS?
Common routes include credential chaining, assuming roles with broad trust, stealing temporary credentials from the metadata service, reading secrets, and creating new keys or policies to escalate.
What is the best defense against lateral movement?
Limit what each identity can reach and how long it can reach it: least privilege, no standing privilege, short-lived credentials and tight trust policies. Add monitoring for unusual identity behavior.
What is the difference between lateral movement and privilege escalation?
Lateral movement is reaching other systems or identities at a similar level. Privilege escalation is gaining a higher level of access. Attackers usually combine them.