Kubeflow: CNCF declares the AI platform mature

Nine years after its first demo, Kubeflow holds the CNCF's top status. An independent security audit with three critical findings preceded the approval.

By Tobias Massow August 24, 2026 5 min read

Kubeflow: CNCF declares the AI platform mature

Kubeflow is now a mature Cloud Native Computing Foundation project. The approval changes little in the code but matters greatly to whether an open AI platform on Kubernetes clears procurement.

Key takeaways

The Foundation promotes Kubeflow. Incubation becomes Graduation: the open-source project is considered mature enough to support production model training, fine-tuning and deployment on Kubernetes.
The approval matters in procurement. “CNCF Graduated” unlocks discussions with procurement teams and risk committees that often stall at “incubating”, according to a contributor with operational experience.
An audit and governance were mandatory. Before approval, the Foundation required an independent security audit, a formal Steering Committee and a commitment to the CNCF Code of Conduct.

Nine years after the hot dog demo

In 2017, a small Google team demonstrated that machine learning could run on Kubernetes. The demo revolved around hot dogs; the name stuck: Kube plus Flow. Since then, a collection of components has become a platform spanning data processing, interactive development, distributed training, fine-tuning and serving.

The Cloud Native Computing Foundation has now granted the project Graduated status. Kubeflow had been incubating since 2023. Graduation signifies technical maturity, robust governance and production use already established at major organizations. As one of the first AI-native projects at this level, Kubeflow provides the foundation for the data and model lifecycle.

According to the Foundation, thousands of contributors across more than a thousand organizations helped build it. Users of its Python packages and subprojects include Bloomberg, NVIDIA, Red Hat, LinkedIn and Spotify. This shows the stack has moved beyond the lab.

Graduation sends a signal to procurement

Deploying models without tying the future to three hyperscalers requires a platform that stays portable. Kubeflow runs wherever Kubernetes runs: public, private or hybrid. Subprojects work individually or together as a community distribution.

The approval changes procurement discussions. Vikas K. Saxena deploys Kubeflow in Australia and New Zealand at organizations often operating without continuous connectivity and under strict oversight. Graduation, he writes, unlocks discussions with procurement teams and risk committees where “incubating” status falls short. Julius von Kohout, a member of the Steering Committee who also works at DHL Data & AI, calls the milestone proof that Kubeflow works as a large enterprise platform.

For European operators, the significance lies in the contract. Where training data resides, who holds the contract and which law applies remain open questions. A vendor-neutral layer on their own cluster at least allows those questions to be raised.

What Kubeflow supports today

The platform is modular. Notebooks support interactive work. Pipelines make workflows repeatable. Trainer handles distributed training and fine-tuning across common frameworks. Katib automates hyperparameter searches. Hub, formerly Model Registry, tracks versions and artifacts. Spark Operator makes Spark jobs as straightforward to run on Kubernetes as other workloads. A central dashboard connects the interfaces users sign into.

Surrounding it are neighbors from the Foundation ecosystem: Prometheus for monitoring, KServe for serving, Feast for features, Kueue for job queues and Istio for secure service-to-service traffic. Anyone already running Kubernetes will recognize the interfaces. Anyone expecting a ready-made studio must first assemble the components.

The roadmap targets what follows conventional training: large language model orchestration, more advanced fine-tuning, broader data processing and agentic workloads. Graduation certifies the current state. The next roadmap slide remains a statement of intent.

The audit preceded approval

The Foundation sets specific requirements for Graduation. Kubeflow delivered an independent security audit, a formal Steering Committee and adherence to the CNCF Code of Conduct. It also holds the Core Infrastructure Initiative’s Best Practices Badge.

The Open Source Technology Improvement Fund published the audit with ADA Logics. In late summer 2025, two engineers examined six subprojects: Katib, Trainer, Spark Operator, Notebooks, Model Registry and Pipelines. The audit covered code review, threat modeling, fuzzing, continuous integration and the supply chain. The result: 14 security-related findings, including three critical ones. The maintainers have addressed the findings. Operators should update to the latest version and verify it within their own networks.

Chris Aniszczyk, the Foundation’s chief technology officer, calls Graduation the moment that establishes Kubeflow as a mature option for enterprise workloads on Kubernetes. Technical work remains the maintainers’ responsibility. The approval simply confirms that this work can be scrutinized under the Foundation’s rules.

What remains unresolved after approval

Graduation simplifies procurement. It does not simplify operations. GPU quotas, queueing, storage classes, workload identities and which data may ever leave the training environment remain tasks for operators. Pinning subprojects individually keeps upgrade windows small. Choosing the full distribution trades convenience for a shared lifecycle.

The community’s invitation is a sober one: meetings are public, and hard-won operational experience feeds upstream development. The benefit is a stack that procurement, operations and oversight can finally discuss together.

Frequently asked questions

What is Kubeflow?

Kubeflow is an open, Kubernetes-native platform for the data and model lifecycle: from processing through interactive development and training to fine-tuning and serving. Subprojects work individually or together as a community distribution.

What does CNCF Graduation mean?

Graduation is the highest maturity level for a Foundation project. Kubeflow meets its requirements through an independent security audit, a formal Steering Committee, a code of conduct and a Best Practices Badge, among other measures. The project had been incubating since 2023.

Does Graduation change my cluster?

Graduation remains a signal for procurement and risk teams. GPU quotas, queues, identities and data flows remain operational responsibilities. The benefits are portability and the ability to pin subprojects independently.

What belongs to the ecosystem?

The core includes Notebooks, Pipelines, Trainer, Katib, Hub, Spark Operator and the central dashboard. Common companions include Prometheus, KServe, Feast, Kueue and Istio. Serving and the feature store remain optional additions.

What did the security audit find?

ADA Logics examined six subprojects and reported 14 security-related findings, including three critical ones. Fuzzing covered Katib, Pipelines, Spark Operator and Model Registry, among others. The maintainers have addressed the findings; keeping deployments current and checking them independently remains essential.