- ERNW researchers argued at Black Hat that a cluster running notebooks, pipelines or scripting features is probably multi-tenant, whether its owners intend that or not.
- They said network policies, pod security standards and admission controls did not stop their exploits, and one vendor treated a reported issue as expected behaviour.
Any company running machine learning tools, scripting features or CI/CD pipelines on Kubernetes is probably running a multi-tenant platform, whether it intends to or not. That was the argument of Lorin Lehawany and Sven Nobis, security analysts at ERNW, in a Black Hat USA 2026 talk. Kubernetes runs applications across many machines. A tenant is any separate user or team sharing the cluster.
What was said
The speakers began with a test. If your cluster runs machine learning workloads, applications with scripting, or CI/CD pipelines (automated chains that build and ship code), you have tenants. A speaker on the show said "you're most probably using multi-tenant platform and yes whether intended or not". Network policies, pod security standards and admission controls, they added, did not stop their exploits.
Here is the mechanism in plain words. In namespace-based tenancy, each tenant gets its own namespace, a named section of one shared cluster. The control plane is how the cluster is managed, for example through the Kubernetes API. The data plane is how deployed workloads communicate. The speakers showed exploits on both layers, in three projects. The excerpts name two: Kubeflow and Istio.
One weakness was cross-namespace references, where a resource in one namespace points to a resource in another. In Kubeflow, a machine learning platform, deploying a notebook also makes its controller create a virtual service. The speakers said a tenant can control certain aspects of it and use that to try to harm other tenants. On Istio, a speaker said: "They trust that Istio provides the right user and Istio does the authentication." The excerpts do not show the full attack chain, so we do not explain it here.
Why it matters
Our reading: the risk is hard to see on a diagram. A leader who pictures tenants as internal teams may never count a notebook feature, a pipeline or a customer-facing scripting box. Each lets someone run code or shape resources inside your cluster. On this evidence, passing a check for network policies and admission controls does not show that tenants are isolated.
The other side
The speakers drew a limit themselves. When you give your own teams a namespace, you share a level of trust and can find and stop an attack, so they called that case not a big risk. The bigger risk, they said, comes from customers and internet users with indirect access.
Fixes are not guaranteed either. They reported each issue to the vendor and hoped for a fix, but said that is not always possible. A speaker said "Istio maintainers see this as a expected behavior as it is a purposeful user experience trade-off." A fix there would have been a breaking change. The excerpts give no count of affected clusters, and the fuller detail sits in a white paper we have not seen.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- Black Hat: Black Hat USA 2026 | Breaking Multi-Tenancy Over and Over, and What We Can Learn From This (2026-11-20)





