Skip to content

Inventory AI workloads with k8s-aibom

k8s-aibom is an unprivileged Kubernetes controller that generates a CycloneDX 1.6 ML-BOM for each AI workload actually running in a cluster — the served model, the runtime, and the container image digests, each attribute carrying an evidence locator and a confidence tier. KubeAI can install it as an optional subchart.

It observes through the Kubernetes API only: no sidecars, no privileged DaemonSet, no changes to the model server pods KubeAI creates.

Enable the subchart

helm upgrade --install kubeai kubeai/kubeai \
  --set k8s-aibom.enabled=true \
  --reuse-values

Opt in the namespace

Inventory is namespace-opt-in by design; nothing is recorded until a namespace is labeled:

kubectl label namespace <kubeai-namespace> aibom.k8saibom.dev/enabled=true

Read the inventory

Each supported workload gets an AIBOM resource. Today that covers Deployments, StatefulSets and the operator CRDs listed in the k8s-aibom README. KubeAI's own Model resources, whose Pods KubeAI creates directly, are not yet covered: that is tracked in GoogleCloudPlatform/k8s-aibom#172 and arrives with k8s-aibom v1.7. Until then, enabling the subchart inventories the other AI workloads in the opted-in namespace, not KubeAI-managed models.

kubectl get aiboms -n <kubeai-namespace>

The kubectl aibom plugin (kubectl krew install aibom) summarizes them:

kubectl aibom summary -n <kubeai-namespace>

Once Model support lands, KubeAI's vLLM and Ollama model servers will be reported with the runtime and model identity declared from the Model spec and image digests from the running Pods. Documents can be shipped to external sinks (object storage, webhook) for audit retention — see the k8s-aibom documentation.