Auditing Kubernetes Cluster Rbac

Auditing Kubernetes Cluster RBAC

When to Use

• When performing security assessments of Kubernetes clusters (EKS, GKE, AKS, or self-managed)
• When validating that RBAC policies enforce least privilege for users and service accounts
• When investigating potential lateral movement or privilege escalation within a Kubernetes cluster
• When compliance audits require documentation of access controls and permissions
• When onboarding new teams to a shared cluster and defining appropriate RBAC policies

Do not use for network policy auditing (use Cilium or Calico network policy tools), for container image scanning (use Trivy or Grype), or for runtime security monitoring (use Falco or Sysdig Secure).

Detection Gaps & Validation

• escalate and bind are the privesc verbs people forget: a Role with bind/escalate on clusterroles/roles lets a subject grant themselves cluster-admin. Also flag impersonate (users/groups/serviceaccounts) and create on serviceaccounts/token (TokenRequest API → mint tokens for any SA).
• Aggregated ClusterRoles hide rules: roles labeled rbac.authorization.k8s.io/aggregate-to-edit/admin silently absorb rules from other ClusterRoles, so a role's printed rules: may understate effective access. Resolve the aggregated result, not the source object.
• RBAC isn't the whole authz story: the system:masters group (mapped via client-cert CN or EKS aws-auth) bypasses RBAC entirely and appears in no binding; cloud IAM→RBAC mappings (EKS aws-auth ConfigMap, GKE IAM roles) grant access kubectl get clusterrolebindings won't show. rbac-tool who-can only evaluates the RBAC authorizer, not webhook/node authorizers.
• Workload-path escalation: create pods + hostPath/privileged/hostPID is effectively node/root access even without secret verbs; SA token automounting lets a compromised pod inherit those rights.
• Validate effective access: confirm with kubectl auth can-i --list --as=system:serviceaccount:<ns>:<sa> (and --as-group); on EKS reconcile against the aws-auth ConfigMap. Note legacy SA token Secrets vs short-lived bound tokens (default in 1.24+) when assessing exposure.

Prerequisites

• kubectl configured with cluster-admin or equivalent read permissions to the target cluster
• rbac-tool installed (kubectl krew install rbac-tool or binary from GitHub)
• KubiScan installed (pip install kubiscan)
• Kubeaudit installed (brew install kubeaudit or from GitHub releases)
• Access to the cluster's audit logs for correlating RBAC findings with actual API access

Workflow

Step 1: Enumerate ClusterRoles and Roles with Dangerous Permissions

Identify roles with wildcard permissions, secret access, pod exec, or escalation capabilities.

# List all ClusterRoles with wildcard verb access
kubectl get clusterroles -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for role in data['items']:
    name = role['metadata']['name']
    for rule in role.get('rules', []):
        verbs = rule.get('verbs', [])
        resources = rule.get('resources', [])
        if '*' in verbs or '*' in resources:
            print(f'ClusterRole: {name}')
            print(f'  Verbs: {verbs}')
            print(f'  Resources: {resources}')
            print(f'  API Groups: {rule.get(\"apiGroups\", [])}')
            print()
"

# Find roles that can read secrets
kubectl get clusterroles -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for role in data['items']:
    name = role['metadata']['name']
    for rule in role.get('rules', []):
        resources = rule.get('resources', [])
        verbs = rule.get('verbs', [])
        if ('secrets' in resources or '*' in resources) and ('get' in verbs or 'list' in verbs or '*' in verbs):
            if not name.startswith('system:'):
                print(f'ClusterRole: {name} -> can access secrets (verbs: {verbs})')
"

# Find roles with pod/exec permissions (container escape risk)
kubectl get clusterroles -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for role in data['items']:
    name = role['metadata']['name']
    for rule in role.get('rules', []):
        resources = rule.get('resources', [])
        if 'pods/exec' in resources or 'pods/*' in resources:
            print(f'ClusterRole: {name} -> has pods/exec access')
"

Step 2: Audit ClusterRoleBindings and RoleBindings

Review bindings to identify who has elevated access and detect overly broad group assignments.

# List all ClusterRoleBindings with the subjects
kubectl get clusterrolebindings -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for binding in data['items']:
    name = binding['metadata']['name']
    role = binding['roleRef']['name']
    subjects = binding.get('subjects', [])
    for subject in subjects:
        kind = subject.get('kind', '')
        subj_name = subject.get('name', '')
        ns = subject.get('namespace', 'cluster-wide')
        print(f'{name} -> Role: {role} | {kind}: {subj_name} ({ns})')
" | sort

# Find bindings to cluster-admin
kubectl get clusterrolebindings -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for binding in data['items']:
    if binding['roleRef']['name'] == 'cluster-admin':
        print(f\"Binding: {binding['metadata']['name']}\")
        for subject in binding.get('subjects', []):
            print(f\"  {subject.get('kind')}: {subject.get('name')} (ns: {subject.get('namespace', 'N/A')})\")
"

# Find bindings granting access to all authenticated users
kubectl get clusterrolebindings -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for binding in data['items']:
    for subject in binding.get('subjects', []):
        if subject.get('name') in ['system:authenticated', 'system:unauthenticated']:
            print(f\"WARNING: {binding['metadata']['name']} grants {binding['roleRef']['name']} to {subject['name']}\")
"

Step 3: Scan with rbac-tool for Comprehensive Analysis

Use rbac-tool for automated RBAC analysis including who-can queries and policy generation.

# Who can get secrets across all namespaces
kubectl rbac-tool who-can get secrets

# Who can create pods (potential for container escape)
kubectl rbac-tool who-can create pods

# Who can exec into pods
kubectl rbac-tool who-can create pods/exec

# Who can escalate privileges (bind/escalate verbs)
kubectl rbac-tool who-can bind clusterroles
kubectl rbac-tool who-can escalate clusterroles

# Generate RBAC policy report
kubectl rbac-tool analysis

# Visualize RBAC relationships
kubectl rbac-tool viz --outformat dot > rbac-graph.dot
dot -Tpng rbac-graph.dot -o rbac-graph.png

Step 4: Run KubiScan for Risky Permissions Detection

Use KubiScan to automatically identify risky service accounts, pods, and RBAC configurations.

# Run KubiScan to find risky roles
python3 -m kubiscan -rroles   # List risky Roles
python3 -m kubiscan -rcr      # List risky ClusterRoles
python3 -m kubiscan -rrb      # List risky RoleBindings
python3 -m kubiscan -rcrb     # List risky ClusterRoleBindings

# Find risky service accounts
python3 -m kubiscan -rs       # Risky service accounts

# Find pods running with risky service accounts
python3 -m kubiscan -rp       # Risky pods

# Check for privilege escalation paths
python3 -m kubiscan -pe       # Privilege escalation vectors

# Generate full report
python3 -m kubiscan -a        # All checks

Step 5: Audit Service Account Token Mounting and Usage

Check for unnecessary service account token mounts that could enable lateral movement from compromised pods.

# Find pods with automounted service account tokens
kubectl get pods --all-namespaces -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for pod in data['items']:
    name = pod['metadata']['name']
    ns = pod['metadata']['namespace']
    sa = pod['spec'].get('serviceAccountName', 'default')
    automount = pod['spec'].get('automountServiceAccountToken', True)
    if automount and sa != 'default':
        print(f'{ns}/{name} -> SA: {sa} (token auto-mounted)')
"

# Find service accounts with non-default token secrets
kubectl get serviceaccounts --all-namespaces -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for sa in data['items']:
    name = sa['metadata']['name']
    ns = sa['metadata']['namespace']
    secrets = sa.get('secrets', [])
    if name != 'default' and len(secrets) > 0:
        print(f'{ns}/{name}: {len(secrets)} secret(s) bound')
"

# Check for pods running as privileged or with host access
kubectl get pods --all-namespaces -o json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for pod in data['items']:
    name = pod['metadata']['name']
    ns = pod['metadata']['namespace']
    for container in pod['spec'].get('containers', []):
        sc = container.get('securityContext', {})
        if sc.get('privileged', False) or sc.get('runAsUser', 1) == 0:
            print(f'RISK: {ns}/{name}/{container[\"name\"]} - privileged={sc.get(\"privileged\",False)} runAsRoot={sc.get(\"runAsUser\",\"not set\")==0}')
"

Step 6: Run Kubeaudit for RBAC and Security Policy Validation

Execute Kubeaudit for comprehensive security checks including RBAC-related findings.

# Run all kubeaudit checks
kubeaudit all --kubeconfig ~/.kube/config

# Run specific RBAC-related checks
kubeaudit privesc    # Check for allowPrivilegeEscalation
kubeaudit rootfs     # Check for readOnlyRootFilesystem
kubeaudit nonroot    # Check for runAsNonRoot
kubeaudit capabilities  # Check for dangerous capabilities

# Output as JSON for processing
kubeaudit all --kubeconfig ~/.kube/config -f json > kubeaudit-results.json

Key Concepts

Tools & Systems

• kubectl: Primary CLI for querying Kubernetes RBAC resources (roles, bindings, service accounts)
• rbac-tool: kubectl plugin for RBAC analysis including who-can queries, visualization, and policy generation
• KubiScan: Python tool for scanning Kubernetes RBAC for risky permissions and privilege escalation paths
• Kubeaudit: Security auditing tool that checks pods and workloads for security anti-patterns including RBAC issues
• rakkess: kubectl plugin showing access matrix for the current user across all resource types

Common Scenarios

Scenario: Auditing an EKS Cluster Shared by Multiple Development Teams

Context: A shared EKS cluster serves four development teams. RBAC was configured during initial setup but has not been reviewed in 12 months. Teams report being able to access other teams' namespaces.

Approach:

1. List all ClusterRoleBindings to identify bindings granting broad access to authenticated users
2. Run kubectl rbac-tool who-can get secrets to find subjects that can read secrets across namespaces
3. Discover that a ClusterRoleBinding grants edit to system:authenticated, giving all users write access cluster-wide
4. Run KubiScan to identify service accounts with risky permissions and pods running with elevated service accounts
5. Replace the ClusterRoleBinding with namespace-scoped RoleBindings for each team
6. Disable automountServiceAccountToken for workloads that do not need API access
7. Create a NetworkPolicy to isolate namespace traffic between teams

Pitfalls: Removing ClusterRoleBindings can break CI/CD pipelines and operators that rely on cluster-wide access. Always audit which workloads use the bindings before removing them. EKS maps IAM roles to Kubernetes groups via aws-auth ConfigMap, so RBAC changes must be coordinated with IAM role mappings.

Output Format

Kubernetes RBAC Audit Report
===============================
Cluster: production-eks (EKS 1.28)
Audit Date: 2026-02-23
Namespaces: 12

RBAC INVENTORY:
  ClusterRoles: 48 (18 custom, 30 system)
  ClusterRoleBindings: 32 (12 custom, 20 system)
  Roles (namespaced): 24
  RoleBindings (namespaced): 36
  Service Accounts: 67

CRITICAL FINDINGS:
[RBAC-001] ClusterRoleBinding Grants edit to system:authenticated
  Binding: authenticated-edit
  Effect: ALL authenticated users have edit access across ALL namespaces
  Risk: Any user can modify resources in any namespace
  Remediation: Replace with namespace-scoped RoleBindings per team

[RBAC-002] Custom ClusterRole with Wildcard Permissions
  ClusterRole: developer-admin
  Rules: verbs=["*"], resources=["*"], apiGroups=["*"]
  Bindings: 4 users via developer-admin-binding
  Risk: Equivalent to cluster-admin without the name
  Remediation: Scope to specific resources and verbs needed

SUMMARY:
  Principals with cluster-admin: 6 (recommended: <= 3)
  Roles with wildcard permissions: 4
  Service accounts with secret access: 12
  Pods with auto-mounted tokens: 45 / 67
  Privileged containers: 8