Kube, Scale Me One More Time ! Exploiting Autoscalers for Kubernetes Cluster Compromise — Alexandre Hervé, Paul Viossat
Date : 04 juin 2025 à 11:30 — 30 min.
Until now, research into node-to-cluster compromise scenarios on Kubernetes has focused on static clusters, where nodes are not dynamically added or removed. While this is sufficient to describe many of the attacks and challenges involved in securing a Kubernetes cluster, it is not representative of the majority of clusters in production. Indeed, over the past few years, autoscaling has become increasingly popular, leading to the integration of Karpenter, the new industry-standard autoscaling solution, into the Cloud Native Computing Foundation in 2024.
Building on our previous research on node isolation in Kubernetes clusters, we will demonstrate that the lack of a unified node designation between cloud providers and Kubernetes APIs can allow an attacker to fully compromise a cluster, starting with access to a single node and without any additional privileges.
To do so, we will explore the frontier between cloud environments, the Kubernetes API, and machines to understand the interactions between these three worlds. This includes how nodes are created, how they join a cluster, and ultimately, how they are deleted. We will deep dive into two major components of Kubernetes cloud clusters: the cloud controller manager and autoscalers.
As actions speak louder than words, we will apply these findings to compromise EKS (AWS) and GKE (Google Cloud) clusters with Cluster Autoscaler or Karpenter managing autoscaling. Finally, we will discuss how to protect against these attacks.