CVE-2026-18982

CVE-2026-18982 is a Kubernetes RBAC failure disguised as an ML training operator bug. The vulnerability isn't in the operator's code — it's in the architectural decision to attach training job creation permissions to the baseline edit and admin ClusterRoles, which transforms what should be a safe middle ground into a cluster-admin pathway. Here's why this matters more than a typical operator vulnerability: edit and admin roles are mental scaffolding. Platform teams build workflows, security policies, and multi-tenant assumptions around the implicit contract that edit = namespace-scoped damage. The training operator breaks that contract silently. Any user with standard edit permissions in any namespace on the cluster — not just in a dedicated ML namespace — can trigger PodTemplateSpec passthrough that enables service account impersonation and host filesystem access. The blast radius isn't ML workloads; it's every workload sharing the cluster. A platform team that hardened their application namespaces is still exposed if someone provisioned this operator on the same cluster. The root cause is a structural incentive problem. Kubernetes permission aggregation makes edit and admin roles powerful by design. Attaching to those roles means the operator works with zero additional RBAC configuration — less setup friction, fewer support tickets. The path of least resistance became a blast radius amplifier because nobody treated it as a blast radius risk. The training operator almost certainly passed security review because nobody interrogates the edit role — it's furniture. What you should do: First, audit every namespace to identify which ones have training operators installed and check what ClusterRole bindings exist. Second, create dedicated training job roles with explicit PodTemplateSpec validation rather than relying on edit/admin aggregation. Third, review other operators in your cluster for the same pattern — PodTemplateSpec passthrough is a known risky pattern, not a single bug. If this operator was installed with edit aggregation, assume the RBAC mental model for your entire cluster is compromised and treat accordingly until you've completed the role scoping exercise. The fix exists but deployment will require touching permissions that production pipelines depend on. Expect remediation timelines measured in quarters rather than days if you have entrenched training workflows.

Reviewed through automated stages and approved by a human before publication.