/audit-security

The /audit-security skill looks for over-privileged identities and workloads: ServiceAccounts with more access than they use, pods running as root or with host-level escapes, and bindings that grant cluster-wide power where a namespace-scoped role would do.

Run it without arguments for a full sweep, or name a workflow to scope the report.

/audit-security                    # full sweep
/audit-security rbac               # single workflow
/audit-security pods in kube-system

Natural-language scoping (namespaces, label selectors, workload names) is supported on every workflow (see Overview).

---

Workflows

1. RBAC

Checks

• ClusterRoles and Roles granting wildcard verbs or resources (*)
• Bindings to cluster-admin and other high-power built-in roles
• RoleBinding and ClusterRoleBinding referencing subjects that no longer exist

Sources: Kubernetes API.

2. Pod security

Checks

• Containers running as root, with no securityContext, or with allowPrivilegeEscalation: true
• Pods with privileged: true, hostNetwork, hostPID, or hostIPC
• Writable root filesystems and dangerous Linux capabilities (SYS_ADMIN, NET_ADMIN, etc.)
• Missing seccompProfile and workloads failing the Pod Security Standards baseline or restricted profiles

Sources: Kubernetes API, evaluated against the upstream Pod Security Standards.

3. Secrets & ServiceAccount tokens

Checks

• Orphaned secrets with no consumer
• Pods with automountServiceAccountToken enabled whose SA has no RoleBindings
• Long-lived legacy kubernetes.io/service-account-token Secrets still present on the cluster

Sources: Kubernetes API.

---

What the agent is told

Beyond the workflows, the skill briefs the agent on how to report:

• Rank findings by blast radius — cluster-scoped wildcards above namespace-scoped ones, host escapes above missing seccomp profiles.
• Be explicit that RBAC checks are static: they find what Roles grant, not what subjects actually use. Detecting truly unused permissions requires audit-log analysis, which this skill does not do.
• Reference Secret objects by name, namespace, and type only — never read or surface contents.
• Explain why a finding matters in one line (what the privilege enables) rather than only naming the offending verb or flag.
• Hand off to /investigate for a specific workload and /audit-network for mTLS and mesh posture, which sits adjacent to pod security.