Anyone who has to decide how strictly a Kubernetes cluster should police itself ends up choosing between a handful of stances. This piece lines three of them up against each other and says plainly which one fits which situation. The subject is KubernetesSecOps, meaning the routine upkeep of cluster permissions, allowed workloads and network conversations. The hospital angle is purely organisational: think of the rules a large facility sets for who enters the records room, and how those rules get rolled out without stopping the day shift.
The three stances are audit-only, warn-then-enforce and enforce-at-once. They are not products; they are ways to introduce the same admission rules. I will compare them on four criteria: risk of breaking working software, speed to protection, effort for the team and what you learn along the way.
What is being compared
The admission layer is the checkpoint that reads every request to create or change a workload and can accept, reject or rewrite it. Kubernetes ships a Pod Security mechanism with three profiles, and teams add policy engines when they want more control. According to the Ops Atlas, the project itself now leans toward expression-based policies built into the API server where that is feasible, and two widely used external engines behave differently when their webhook is unreachable: Kyverno stops requests by default, while Gatekeeper lets them through by default, as of the Atlas’s October 2026 note. That difference matters later in this comparison.
For a broader walk through the topic, this long Ops Atlas page on KubernetesSecOps covers every layer from access lists to runtime alerts. Here we stay on the rollout choice.
Stance one: audit-only
In audit-only mode the checkpoint records what it would have refused but lets everything through. Nothing breaks, so nobody complains, and that is exactly the problem: nothing is protected either.
- Breakage risk: close to zero.
- Speed to protection: none by itself.
- Team effort: low at first, then growing as the report of violations piles up unread.
- Learning: excellent. You get an honest inventory of which workloads break which rule.
Audit-only is the right first week. It is the wrong sixth month. A common failure is a dashboard full of findings that become part of the scenery.
Stance two: warn-then-enforce
Here the checkpoint accepts the request but prints a warning to whoever submitted it, and the team publishes a date on which the rule becomes binding. The Atlas suggests a warning period of roughly two weeks before the switch, followed by enforcement paired with an expiring register of exceptions.
- Breakage risk: low, because owners see the message while they can still fix things.
- Speed to protection: moderate; you wait out the warning window.
- Team effort: medium. Someone must announce the date, chase owners and track exceptions.
- Learning: good. Real owners tell you which rules are unreasonable.
This is the stance most organisations should default to. It trades a few weeks for goodwill, and goodwill is what keeps a rule in place for years.
Stance three: enforce-at-once
The rule is switched on and rejections start immediately. A new request that violates it is refused. There is one detail worth knowing: enforcement only affects requests that arrive after the switch. Pods that are already running are not evicted, so a cluster can look compliant on paper while carrying older workloads that would fail today.
- Breakage risk: high for deployments and automation that relied on the old behaviour.
- Speed to protection: immediate for anything new.
- Team effort: front-loaded and urgent. Tickets arrive the same afternoon.
- Learning: poor. You find out what breaks by breaking it.
The Atlas names the extreme version of this as an anti-pattern: switching on every sample policy in enforcing mode. Reserve this stance for a brand-new namespace with nothing in it, or for a rule that guards something catastrophic, where the cost of a few failed deployments is clearly smaller than the risk.
Side-by-side summary
- Choose audit-only for discovery, with an end date written down before you begin.
- Choose warn-then-enforce for the bulk of existing workloads and for rules that touch many teams.
- Choose enforce-at-once for new, empty spaces or a short list of non-negotiable rules.
A generic scenario to make the trade-offs concrete
Picture a mid-sized organisation with forty namespaces, each owned by a different application team, and a goal of stopping containers from running with unnecessary powers. This is an illustration, not a report of any real deployment. With audit-only, the platform team learns within days that a good share of namespaces would fail the rule, mostly because of a handful of shared base settings. With warn-then-enforce, those owners receive messages on every deploy and fix the shared settings once, which clears most of the noise in a single stroke. With enforce-at-once, the same shared settings would have caused a wave of rejected releases on day one, and the platform team would have spent that week apologising instead of improving anything.
The example also shows why exceptions need an expiry. A legitimate exception, say for a monitoring agent that needs extra access, is fine. An exception without a review date becomes a permanent hole, and the Atlas’s workflow treats an expiring exception register as part of the enforcement step, not an optional extra.
Two traps that hit every stance
First, rules attach to pods, not to the higher-level objects that create them. A policy that inspects a Deployment but not the pods it spawns can be sidestepped, so test with the object type the rule really guards. Second, waving through one automation identity also waves through any person able to launch workloads under it. Such exemptions look tiny in a config file yet reach surprisingly far.
Both come from reading a policy as a statement of intent instead of testing what it actually blocks. A suite that tries to deploy a deliberately bad workload, and expects refusal, pays for itself the first time somebody edits a rule.
The failure-mode question that changes the comparison
All three stances rely on the checkpoint being reachable. If the webhook that makes decisions is down, the cluster must either refuse requests or let them pass. A fail-closed setup protects you but can block even routine restarts; a fail-open setup keeps things running but silently drops your protection at exactly the wrong moment. The Atlas’s guidance is to pick the behaviour per tier of workload rather than globally, and it lists a blanket “ignore failures” setting among the anti-patterns. Ask your platform team what the current setting is and why.
One more practical note on cost. Warn-then-enforce looks slower on a calendar, yet the total effort is often smaller than the alternatives, because owners fix their own workloads while the date is still ahead of them. Audit-only shifts the work onto a central team that must read every finding, and enforce-at-once shifts it onto whoever is on call. When you present the options to a budget holder, put the person-hours of each stance in the table beside the calendar time, and the middle option usually wins the argument without any persuasion.
Comparing the way you would compare facilities
Practical readers may like a model for structured comparison. A piece from the same site on comparing providers against a fixed list of criteria follows the habit of writing the criteria first and scoring afterwards. The habit transfers neatly: decide your four criteria before you look at any policy engine, or the loudest feature will pick for you.
If your team also supports wider platform work in Japan or elsewhere, the enterprise guide covering DevOps, Kubernetes and SRE gives the surrounding delivery context in which these admission rules sit.
A practical recommendation
Start in audit-only for a week and no longer. Move the broad rules to warn-then-enforce with a calendar date and a named owner for the exception register. Use enforce-at-once only where the blast radius of a mistake is small or the rule is truly critical. Review rejections weekly, and keep the exception list short by giving every entry an expiry. That sequence is not glamorous, but it is the one that survives contact with a busy organisation.
