Seccomp

Every action a container takes (opening files, sending packets, spawning processes) ends up as a syscall into the shared node kernel. The kernel exposes hundreds of syscalls; a typical application uses well under a hundred. The rest are pure attack surface, and kernel vulnerabilities are regularly reachable only through obscure syscalls.

Seccomp (secure computing mode) filters which syscalls a process may make. Kubernetes exposes it per pod or per container through seccompProfile:

Type	Meaning
Unconfined	No filter. The historical default.
RuntimeDefault	The container runtime’s curated profile, blocks 50+ dangerous or obscure syscalls
Localhost	A custom JSON profile from the node’s seccomp directory

Enable RuntimeDefault

This is the highest-value, lowest-effort setting on this page:

apiVersion: v1
kind: Pod
metadata:
  name: seccomp-default
  namespace: lab
spec:
  securityContext:
    seccompProfile:
      type: RuntimeDefault
  containers:
    - name: app
      image: busybox:1.36
      command: ["sleep", "3600"]

Verify the filter is active:

kubectl apply -f seccomp-default.yaml
kubectl exec -n lab seccomp-default -- grep Seccomp /proc/1/status

Expected output:

Seccomp:        2
Seccomp_filters:        1

Mode 2 means a filter is loaded (0 is unconfined). The runtime’s default profile blocks syscalls like mount, reboot, init_module, and ptrace variants, things applications do not do but exploits love.

Write A Custom Profile

Localhost profiles let you go tighter. This profile allows everything except the chmod family:

{
  "defaultAction": "SCMP_ACT_ALLOW",
  "syscalls": [
    {
      "names": ["chmod", "fchmod", "fchmodat"],
      "action": "SCMP_ACT_ERRNO"
    }
  ]
}

The kubelet loads profiles from /var/lib/kubelet/seccomp on each node. On the kind lab cluster, place it there through the node container:

docker exec k8s-security-lab-control-plane mkdir -p /var/lib/kubelet/seccomp
docker cp deny-chmod.json k8s-security-lab-control-plane:/var/lib/kubelet/seccomp/deny-chmod.json

Reference it by path relative to that directory:

apiVersion: v1
kind: Pod
metadata:
  name: seccomp-custom
  namespace: lab
spec:
  securityContext:
    seccompProfile:
      type: Localhost
      localhostProfile: deny-chmod.json
  containers:
    - name: app
      image: busybox:1.36
      command: ["sleep", "3600"]

kubectl apply -f seccomp-custom.yaml
kubectl exec -n lab seccomp-custom -- chmod 777 /tmp

Expected output:

chmod: /tmp: Operation not permitted
command terminated with exit code 1

The syscall was refused by the kernel before it did anything.

Real allowlist profiles (default deny, explicit allow list) are stronger but must be generated from observation. The Security Profiles Operator can record what a workload actually calls and distribute the resulting profile across nodes.

Clean Up

kubectl delete pod seccomp-default seccomp-custom -n lab
docker exec k8s-security-lab-control-plane rm /var/lib/kubelet/seccomp/deny-chmod.json

Practical Guidance

1. Set seccompProfile: RuntimeDefault on every pod today; breakage is rare and the win is large.
2. Enforce it via the restricted Pod Security Standard, which requires exactly this field.
3. Reserve custom Localhost profiles for high-risk, well-understood workloads; they must be distributed to every node the pod can schedule on.
4. Generate allowlist profiles from recordings rather than writing them by hand.
5. When a workload crashes under a profile with errors like Operation not permitted, compare with Unconfined in dev first, because that isolates seccomp as the cause before you start editing profiles.