airom k8s enumerates the container images referenced by Kubernetes workloads, then scans each unique image.
Live-cluster mode is not wired up. airom k8s without --manifests does not talk to a cluster — it fails immediately with a clear error. Offline manifest mode (--manifests <dir>) is what works today, and it is what this page documents.

​
Offline manifest mode

Enumerate images from rendered manifests
AIROM walks the directory, parses every .yaml/.yml file (multi-document streams included), extracts the container images from workload specs, dedupes them, and sorts them. Rendered YAML and helm template output both work.
1

Render your manifests

Helm, for example
2

Enumerate and scan

​
Which workloads are read

Pod, Deployment, StatefulSet, DaemonSet, ReplicaSet, Job, CronJob. Any other kind is skipped. Within a workload, every container position is read: containers, initContainers, and ephemeralContainers. CronJob’s nested job template is handled.

​
Enumeration output

Each image is reported once, sorted by reference, with the number of distinct workloads that reference it. AIROM then scans each image in turn and emits an AIBOM per image, prefixed with an === image <ref> === header. A file that cannot be read, or a document that fails to parse, degrades to an Unknown and the walk continues — one bad manifest never kills the run.

​
The catch: image refs must be resolvable locally

Manifest enumeration gives you image references (ollama/ollama:latest). Scanning those references requires resolving the image — and, as documented in /scanning/container-images, live registry and daemon pulls are not wired up in this build.
A registry reference pulled out of a manifest cannot be fetched today. AIROM logs a warning for each image it could not scan and moves on. In practice the enumeration is the reliable part of k8s --manifests; scanning succeeds only for references that resolve to a local image archive or OCI layout path.
The dependable workflow today is two steps:
1. Enumerate what your cluster runs
2. Scan each image from a saved archive

​
Flags

​
string
Directory of manifest YAML to enumerate workload images from. Required today — without it, k8s reports that live-cluster scanning is unavailable.
​
string
Restrict to one namespace. Accepted and parsed, but it applies to live-cluster enumeration, which is not wired up. In manifest mode it does not filter — every workload document in the directory is read regardless of its namespace.
​
bool
default:"false"
All namespaces. Same status as --namespace: accepted, but with no effect in manifest mode, which already reads every document in the directory.
​
bool
default:"false"
Scan images concurrently. Accepted and parsed, but not yet honored — images are scanned serially today, whether or not you pass it.
The optional [context] positional selects a kube context. It is accepted, but it only means something in live-cluster mode, which is not wired up.
--parallel (the global worker-count flag) is unaffected by any of this. It controls detector workers within a single image scan and works normally.

​
CI gating across images

--fail-on is evaluated per image. The gate trips if any scanned image matches the policy, and the run then exits with --exit-code (default 1).
Fail if any workload image ships a hosted LLM reference
Images that could not be scanned are reported as warnings — they do not trip the gate. Given the resolution limitation above, do not treat a green k8s --manifests gate as evidence that your workload images are clean. See /reference/cli for the full --fail-on grammar and exit-code contract.