A few weeks ago, once a long day of flying had ended, a colleague who runs the technology side of a flying operation pulled me aside. “Everyone keeps saying our booking system runs on Kubernetes and needs KubernetesSecOps,” she said. “I nod along. Can you tell me what it means, without the jargon?” We ended up talking for an hour. What follows is that conversation, rearranged as the questions she actually asked, with the answers I would give again. The reference material is the Ops Atlas, which has a long entry on the topic.
What does KubernetesSecOps actually mean?
It is shorthand for the continuing work of keeping a Kubernetes cluster safe. Kubernetes is the software that starts, stops and moves the small packaged programs that make up modern apps. KubernetesSecOps asks five questions about it: who is allowed to change things, what is allowed to run, what is allowed to talk to what, where passwords and keys are kept, and how we would notice if somebody misbehaved. The Atlas gives the subject a dedicated page, and its KubernetesSecOps article is the one I pointed her to for the long version.
She compared it to a safety management routine in a flight school: nobody expects a single heroic act, just habits that run every week.
Is it a job, a certificate, or a product?
None of the three. The Atlas says the name is an unusual label for what the wider world calls Kubernetes security. People take an exam called the CKS to prove the skill, after a more general administrator exam. If a vendor tells you that buying a product makes you “KubernetesSecOps compliant”, remember there is no such standard.
Our cluster works fine. Why worry?
Because “works fine” and “is protected” are different statements. I told her about the three defaults that surprised me most when I first read the Atlas. Passwords and keys stored as Secrets are held unencrypted in the cluster’s database unless somebody switches encryption on. Every program can reach every other program until a network rule says otherwise. And on a cluster you run yourself, nothing is written to an activity log until you provide a logging policy. She laughed and said it sounded like an aircraft delivered with the doors unlocked and the logbook blank.
Does hiding a password in base64 count as protection?
No, and this one comes up constantly. Base64 is a way of writing data using only ordinary characters. Anyone can reverse it in seconds. Real protection means encryption at rest, and the Atlas recommends the current key-management approach, and it adds that the older approach lost support in release 1.28 and was disabled by default one release later.
Who should be allowed to change the cluster?
Fewer people than you think, with narrower powers than you think. The Atlas lists permissions that deserve extra care: reading Secrets, starting new workloads, running commands inside live ones, impersonating other users and approving certificate requests. A pipeline that deploys software should hold only what it needs, never full administrator rights. And nobody should be put in the special group that skips permission checks.
We talked about how this mirrors a hangar: the person who fuels the aircraft does not need the key to the parts cage, and the stores clerk does not need to sign off a flight release.
What is “admission”, and why does it matter?
Admission is the checkpoint a request passes after permission has been granted and before anything is stored. Think of it as the dispatcher checking the load sheet. It matters because people can deploy with command-line tools that bypass your delivery pipeline, so a check only in the pipeline can be walked around. The Atlas therefore recommends enforcing at the cluster door as well.
One caution she wrote down: a new rule should begin in observation mode, move to warnings for a fortnight against an announced date, and only then start refusing. It is the same courtesy as publishing a new procedure before the next check ride.
Are namespaces walls between teams?
No. A namespace is a label with permissions attached. Programs placed on one machine all lean on a single operating-system kernel, so a serious flaw in one can reach its neighbours. If two groups truly must be kept apart, give them separate machines or separate clusters. Current Kubernetes guidance, the Atlas notes, thinks in lifecycle phases (develop, distribute, deploy, runtime) in place of the older four-layer picture.
How would we even notice an attack?
Three sources help: the audit log, the stream of refused requests, and runtime detection that watches running programs for odd behaviour. The Atlas includes a sample month with 1,240 alerts, yet merely 62 deserved a response, around 5 percent. Filtering out routine, already-audited engineer sessions left 140 alerts with 58 actionable. She asked whether a noisy alarm is worse than none. Often, I said, yes, because people learn to ignore it.
When something does look suspicious, resist the urge to delete the pod instantly. Doing so destroys evidence and does not cancel any keys the program held. A calmer response is to change its labels so existing network rules no longer cover it, then investigate.
How often should the reviews happen?
The Atlas proposes a rhythm: refusals and alerts skimmed weekly, the exemption list read monthly, permissions and a benchmark scan redone quarterly, and one rehearsal each year with whoever handles incident triage. The maximum gap between full benchmark and permission reviews is given as two years, which she compared to a regulatory interval that nobody aims to use up.
Do we need special software to begin?
Not to begin. The first stages of the Atlas plan use features that already ship with Kubernetes: network policies, pod safety labels, audit logging and role reviews. Specialist tools come later, for detection, signature checks and policy at scale. A small team can learn a great deal with a lab cluster and a notebook. In flying terms, you learn stalls in a trainer before you ever consider the avionics upgrade.
What if a rule breaks something important?
This is why the warning phase exists. Because refusals are previewed first, the breakage shows up in logs and messages while nothing is yet blocked. If a rule still causes trouble after enforcement, use an exception that names an owner, a reason and a date, and then fix the underlying workload. The register of such exceptions is the quiet hero of the whole approach.
Where do we start on Monday?
I gave her three steps, in order. First, make a list of every namespace and write beside each whether a network rule exists. Second, switch on audit logging and confirm something is recorded. Third, find out who holds broad administrator rights and write down why. That is the “see” stage of the Atlas 30-60-90 plan. Next comes “guard”, with baseline rules enforced, a deny-all pilot, encryption and alerts routed to a real person. After that, “verify”, with the stricter profile on new namespaces and rehearsals.
She finished her coffee and said a summary that beats most of the slide decks I have seen: “So it is mostly looking, writing things down and not trusting the defaults.” Close enough. The rest is doing it on a calendar, one small habit at a time, until the cluster behaves the way the paperwork claims.
Anything people get wrong about all this?
Yes, four things come up repeatedly. People assume a clean scan result proves a clean cluster, when scanners can miss some permission routes. They think blocking every image with a flagged vulnerability at the gate is strict, when the Atlas lists it as a pattern that causes more harm than it prevents. They believe an exemption is harmless, when one granted to a controller’s identity extends to anyone who can create that workload. And they treat the work as finished after one successful audit. Safety culture, in aviation and in clusters, is the habit of looking again.
