k8scout

Beta — A single-binary Kubernetes attack path engine for authorized security assessments.

Drop it into a compromised pod, run it, and get a map of every realistic escalation path — from your current foothold to cluster-admin, node access, secret theft, and cloud IAM roles.

---

The idea

You have RCE in a Kubernetes pod. Now what?

k8scout answers that question. It automatically discovers what the compromised pod's service account can do, maps out the RBAC graph, and traces multi-step attack paths from your exact foothold to high-value targets.

It works in three modes:

• Offensive mode (default) — run from inside a compromised pod. Discovers your identity, permissions, and all reachable escalation paths from your current position.
• Reviewer mode (--reviewer-mode) — run with a read-only SA to audit the full cluster attack surface for all identities.
• Recon mode (--recon) — a fast, low-footprint look at just your identity, effective permissions, and which resources you can touch. No graph or pathfinding. Add --bruteforce-ns to discover namespaces you can't list cluster-wide.

---

Demo

---

What it finds

k8scout builds a weighted permission graph and runs Dijkstra-based pathfinding to discover realistic multi-step attack chains:

• Pod to cluster-admin — through RBAC bindings, workload mutation, or CRB creation
• Container escape to node — via privileged containers, hostPID, hostNetwork, dangerous capabilities, or hostPath mounts
• Lateral movement — exec into other pods, steal their SA tokens, pivot through their permissions
• Secret and credential theft — mounted SA tokens, secrets, configmaps with leaked credentials, orphaned tokens
• Cloud IAM escalation — IRSA (AWS), GKE Workload Identity, Azure Workload Identity, projected token audience abuse
• Impersonation chains — SA-to-SA takeover through impersonation permissions
• Workload mutation — patch a deployment to change its SA, then inherit that SA's permissions
• Webhook injection — mutating webhook control to inject into future workloads
• GitOps operator abuse — ArgoCD, Flux, External Secrets, and Vault operator privilege escalation
• Misconfiguration detection — dangling bindings, wildcard verbs, automounted tokens, plaintext secrets in env vars

Every finding includes MITRE ATT&CK technique IDs, a risk score, and step-by-step attack path with the actual graph nodes involved. 50 detection rules in total.

---

Quick start

From a compromised pod (primary use case)

# Copy the binary into the pod
kubectl cp k8scout-linux-amd64 <ns>/<pod>:/tmp/k8scout

# Run it
kubectl exec -it <ns>/<pod> -- chmod +x /tmp/k8scout
kubectl exec -it <ns>/<pod> -- /tmp/k8scout --out /tmp/result.json

# Pull the results
kubectl cp <ns>/<pod>:/tmp/result.json ./result.json

# This flow works exactly like this in you're revshell!

The binary auto-detects it's running in-cluster, identifies the pod and SA, and starts pathfinding from your exact foothold.

Intended use: Penetration testing engagements, red team operations, internal security reviews, and cluster hardening audits. Always obtain proper authorization before running against any cluster.

From your local machine

# Uses ~/.kube/config or $KUBECONFIG
k8scout --all-namespaces --out result.json

# Target a single namespace
k8scout --namespace production --out result.json

Recon mode (quick permission & resource check)

# Who am I, what can I do, and what resources can I access?
k8scout --recon

# Discover namespaces you can't list cluster-wide (built-in wordlist)
k8scout --recon --bruteforce-ns

# Bruteforce with your own namespace wordlist (one name per line)
k8scout --recon --bruteforce-ns --ns-wordlist ./namespaces.txt

Recon skips graph building, inference, and AI — it's the fastest way to answer "what does this token actually have?". It reports the current identity, its effective permissions (SelfSubjectRulesReview), and a per-resource capability matrix (SelfSubjectAccessReview). --bruteforce-ns confirms a namespace exists by reading well-known objects every namespace has (the namespace object, its default ServiceAccount, or the kube-root-ca.crt ConfigMap) — useful when the identity can't list namespaces but may still access specific ones.

Reviewer mode (full cluster audit)

# Deploy the read-only RBAC and job
kubectl apply -f deploy/rbac.yaml
kubectl apply -f deploy/job.yaml

# Or run directly with reviewer permissions
k8scout --reviewer-mode --all-namespaces --out result.json

With AI narrative

export OPENAI_API_KEY="sk-..."
k8scout --all-namespaces --out result.json

---

Installation

Pre-built binaries

Download from the Releases page. All binaries are statically linked with no dependencies.

Binary	Platform
k8scout-linux-amd64	Linux x86-64
k8scout-linux-arm64	Linux ARM64
k8scout-darwin-amd64	macOS Intel
k8scout-darwin-arm64	macOS Apple Silicon

Build from source

Requires Go 1.22+.

git clone https://github.com/hac01/k8scout
cd k8scout
make build          # native binary
make build-linux    # static Linux amd64 binary
make build-all      # all four release targets

---

How it works

k8scout (running inside compromised pod)
 |
 ├── 1. Detect foothold
 │     Pod name (HOSTNAME), SA (TokenReview), Node (downward API)
 │
 ├── 2. Discover permissions
 │     SSRR per namespace + ~30 concurrent SSAR spot-checks
 │     (always permitted, no RBAC needed)
 │
 ├── 3. Enumerate cluster objects (graceful degradation if denied)
 │     Namespaces, RBAC, Workloads, Pods, Secrets, Nodes, Webhooks, CRDs
 │
 ├── 4. Build attack graph
 │     Nodes: pods, SAs, roles, bindings, secrets, workloads, nodes, cloud identities
 │     Edges: runs_as, mounts, authenticates_as, can_exec, can_patch,
 │            can_impersonate, runs_on, assumes_cloud_role, granted_by, ...
 │
 ├── 5. Find attack paths (Dijkstra from foothold to high-value targets)
 │     Weighted by attacker effort — cheapest (most realistic) paths first
 │     Targets: cluster-admin, nodes, SA tokens, cloud IAM, privileged workloads
 │
 ├── 6. Run inference rules (50 rules with MITRE ATT&CK mapping)
 │
 ├── 7. Optional: AI risk narrative (GPT-4o)
 │
 └── 8. Output: text summary + JSON report

Even with a minimal SA that can't list pods or RBAC objects, the tool synthesizes the foothold graph from identity data alone and discovers what's reachable through SSRR/SSAR permissions.

---

Attack graph visualization

Load the JSON report into web/graph.html in any browser (drag-and-drop, no server needed).