How to use the Kubernetes Cluster Attack Techniques agent skill

What the skill makes the agent do

An agent skill is a folder containing a SKILL.md file with instructions that the agent reads when a request matches the skill's description. This one contains a structured workflow for Kubernetes: determine the starting position, enumerate permissions and resources, identify escalation vectors, chain misconfigurations, and then look at secrets, other namespaces and cloud metadata.

The description lists the attack areas it covers: service account token theft and impersonation, RBAC misconfiguration including wildcard permissions and role binding escalation, etcd access for secret extraction, kubelet API abuse on ports 10250 and 10255, pod escape through hostPID, hostNetwork and hostPath, secrets enumeration, admission controller and network policy bypass, cloud metadata access on EKS, GKE and AKS, CRD and operator persistence, and node compromise through DaemonSets. It maps these to MITRE ATT&CK techniques T1609, T1610 and T1613.

The skill names the tools it expects the agent to use: kubectl, kube-hunter, peirates, kubeaudit, kdigger and kubeletctl. Where kubectl is missing, the SKILL.md tells the agent to fall back to curl with the service account token.

Install the skill on your machine

Download the ZIP for this skill from this page. Unzip it so the offensive-k8s-attacks folder, containing SKILL.md, sits in your personal skills directory at ~/.claude/skills/. To limit it to one project instead, place the folder in that project's .claude/skills/ directory.

If you use the Claude apps rather than a local setup, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal install: ~/.claude/skills/offensive-k8s-attacks/
• Project install: .claude/skills/offensive-k8s-attacks/ inside your project
• Claude apps: Settings → Capabilities → Skills

Trigger the skill with natural requests

You do not need to name the skill. The agent loads it when your request matches the description, which mentions assessing Kubernetes clusters, attacking from within a compromised pod, exploiting RBAC or kubelet misconfigurations, and cloud-native lateral movement. State the scope and your starting position in the request so the agent knows what it has.

Check the first step before going further. The skill begins by establishing your position, so tell the agent what you already have.

• I have a shell in a pod in an authorised test namespace. Work out what my service account can do.
• Use the offensive-k8s-attacks skill to check my RBAC permissions from this kubeconfig, which is part of a signed-off lab cluster.
• Audit this lab cluster's RBAC roles for wildcard permissions and tell me which bindings are risky.
• I am in a deliberately vulnerable training cluster. Which kubelet misconfigurations are exposed on the nodes?

What you need before you start

kubectl is optional. Without it the skill falls back to curl with the service account token. You also need Kubernetes cluster access that you are authorised to test, such as a pod shell, a kubeconfig file or an API server endpoint.

The third-party tools kube-hunter, peirates, kubeaudit, kdigger and kubeletctl are optional. The skill references them, so install only the ones your engagement needs. The skill runs scripts and uses the network, so review what the agent executes in your environment.

• kubectl (optional)
• Cluster access you are authorised to test
• Optional: kube-hunter, peirates, kubeaudit, kdigger, kubeletctl

Safety, scope and rollback

This skill performs active enumeration and exploitation. The SKILL.md itself frames its objective as escalating privileges, moving laterally and compromising the cluster or underlying cloud infrastructure, so it must only be used against systems you have written authorisation to test. Attacking clusters or cloud accounts without that authorisation is illegal and harms the owners.

Do not point it at production clusters without a rollback and scope plan. Enumeration and exploitation can disrupt workloads or trigger alerts. The skill is an offensive playbook and does not include hardening steps, so use it to find the weaknesses you then fix.

• Confirm written scope and authorisation before the first request.
• Prefer a lab or deliberately vulnerable cluster for learning.
• Set a rollback plan for any change the agent proposes, such as creating roles or bindings.
• Review every script and network call before running it in a shared environment.

Tips for better results

Give the agent the least information it needs to start, then add context as findings come in. Ask it to report each finding with the permission or misconfiguration behind it so you can map it to a fix. Because the skill is organised as phases, you can ask for only one phase, such as enumeration, and stop there.

• Start with enumeration only and review results before escalation.
• Name the namespace and the identity you are testing to keep scope tight.
• Ask for the MITRE ATT&CK technique behind each finding for your report.
• Use a lab cluster first to see how the agent behaves.