KubeCon North America 2026 Wrap-Up: AI, Application Security, and Virtualization

Christian Hernandez, Katie Meinders, Marino Wijay

KubeCon North America 2026 Wrap-Up: AI, Application Security, and Virtualization

Running AI in production brought familiar Kubernetes questions into focus. How would platform teams scale these workloads, recover from failures, and keep automated actions within defined boundaries?

Those types of problems were recurring themes at this year's KubeCon + CloudNativeCon North America 2026 in Salt Lake City, UT. AI infrastructure and agents dominated the conference, alongside application security and running virtual machines on Kubernetes. Across those topics, the emphasis was on the operational work required to run these systems reliably.

This blog looks back at those themes and what they suggested about the cloud native community's priorities. We’ll also take a look at what transpired during the week.

Cilium Everywhere

CiliumCon took place on Monday, November 9, with a full morning of sessions on Cilium, Tetragon, and eBPF. The program mixed end-user talks from ESnet, OpenAI, and Datadog with sessions from Cilium maintainers and contributors.

Several sessions focused on multi-cluster networking and IPv6. ESnet shared how they run ClusterMesh with native routing, BGP, and multi-pool IPAM on an IPv6-only network, while contributors from Microsoft looked at what makes multi-cluster networking practical at cloud scale. Isovalent’s Marino Wijay made the case for IPv6 as AI clusters run out of private IPv4 space.

Datadog walked through migrating hundreds of thousands of nodes from veth to netkit with rollback at every step, and OpenAI used eBPF probes to trace rare packet loss in inference workloads. On security, Isovalent’s Liz Rice gave a keynote on using Tetragon for guardrails and mitigations that can be deployed without downtime. Robin Gögge of Isovalent and Quang Nguyen of Microsoft explained how sidecarless mTLS works.

Following CiliumCon, contributors met for the Cilium Developer Summit to discuss project health, including how to encourage productive collaboration as generative tools change how people contribute. Other sessions covered the paths forward for datapath plugins, host firewall behavior with IPsec, IPv6 migration, ztunnel policy, Dynamic Resource Allocation, and how Cilium and Kubernetes network policies interact.

inline-image-0

Cilium was on the main conference schedule too. Isovalent’s Christine Kim and Katie Meinders, and Hadrien Patte of Datadog presented the Maintainer Track session on Cilium's evolution, from Gateway API features to new eBPF datapath plugins. GEICO gave two talks on pairing Cilium with Istio and SPIFFE/SPIRE for network security and multi-cloud platforms, and Daniel Borkmann and Aaron Campbell showed how new Dynamic Resource Allocation drivers bring zero-copy networking to Pods through netkit.

Between sessions, the Cilium booth in the Project Pavilion was home to plenty of good conversations with users and those looking to get started.

eBee Quest

We continued the momentum after CiliumCon by getting hands on at our eBee Quest! With our experts on hand to help, participants spent the evening working through labs in the Cottonwood Room at the Hilton Salt Lake City Center. The attendees completed 28 labs in a friendly competition, with some teaming up to work through the challenges. The winner had never taken an Isovalent lab before. He walked away with a big Star Wars LEGO set and was keen to keep exploring more of the labs.

It was an amazing opportunity to connect with each other in a relaxed environment!

AI Workloads

AI was the most prominent theme throughout the conference. The topics covered the infrastructure behind training and inference, including GPU scheduling and distributed inference. Sharing accelerators and improving utilization also received attention, connecting performance questions to the cost of operating these environments.

Daniel Borkman, Aaron Campbell, and Bernadro Soares’ talk on DRA went over some of this.

Reliability remained part of that discussion. Running AI workloads on large Kubernetes clusters brought attention to cascading failures and recovery. Scheduling also involved more than finding spare capacity. Training jobs needed suitable, connected hardware, making placement and coordination part of the operational problem.

Agents extended the conversation beyond infrastructure. They appeared across multiple tracks, including platform engineering and operations. Troubleshooting, scaling, and responding to alerts were among the uses explored, along with the orchestration and failure recovery required to support them.

Giving software the ability to act also raised questions about its authority. Workload isolation and sandboxing appeared alongside restrictions on credentials, network access, and tool use. Human approval for higher-risk actions was another consideration. The recurring question was how to keep an agent's actions bounded and attributable, with enough visibility to understand what happened and recover from mistakes.

Security and reliability were closely connected throughout these topics. An agent's permissions affected what it could change, while retries and recovery affected what happened when an action failed. Production readiness required attention to both.

Application Security

AI also appeared on the vulnerability/discovery side of the conference. Discussions covered its ability to uncover real zero-days, alongside the work required to validate the results. A plausible report still needed investigation before maintainers could establish its impact.

The volume and quality of generated reports were recurring concerns. Duplicates and findings outside a project's threat model added to the triage burden. A polished report could still leave maintainers with substantial work to reproduce the issue and decide whether it represented a vulnerability.

Reproduction, impact assessment, disclosure, and fixes all required time and coordination. Faster report generation raised questions about the capacity of those response processes and the evidence needed to make a report actionable.

For operators, the related concern was how to prioritize vulnerability information and reduce exposure. The security topics questioned whether scanning and patching alone were enough as AI influenced both discovery and attacks. Runtime context and protection were part of that discussion, alongside the difficulty of separating urgent risks from a growing volume of findings.

Virtualization

Virtual machines were a substantial secondary theme. KubeVirt featured in discussions about operating VMs alongside containers, with attention to VM lifecycle management and storage. Guest observability and troubleshooting were also part of the operational discussion.

The emphasis was on coexistence and gradual modernization. Traditional applications could remain in VMs while other services ran as containers. Managing that mixed environment brought questions about how Kubernetes operations would extend across different workload types.

Networking was part of that operational picture. Sessions covered assigning network devices and GPUs to VMs for training and inference, connecting virtualization to the AI infrastructure themes. These workloads brought their own placement and connectivity requirements into the discussion.

Portable VM management and broader hypervisor support also appeared. The direction was toward a common Kubernetes-facing management model across different underlying platforms, while accounting for the differences between them.

HiveMind Mingle

Isovalent HiveMind Mingle is a community networking event that brings together people from the Cilium, eBPF and broader cloud-native ecosystems. It's an opportunity for Cilium users, contributors, maintainers and industry practitioners to meet, connect and socialize in a more informal setting. The focus is on community building, strengthening relationships and encouraging conversations across the ecosystem. This year, it was held at Squatters Brewery in Salt Lake City.

Summary

AI and agents were the dominant themes at KubeCon North America 2026, with an emphasis on operating them in production. Application security brought attention to AI-assisted discovery and the demands it placed on maintainers and response teams. Virtualization reflected continued interest in managing VMs and containers together through gradual modernization.

Production readiness was the common thread.

The broader concerns remained familiar to Kubernetes operators. Reliability, security, and observability continued to shape the work, even as AI agents and mixed workloads added new operational questions.

What now? We are working on preparation materials for the new CKNE Certification from the Linux Foundation. Come and join us as we walk you through some helpful tips!

Get CKNE-ready with the Isovalent CKNE Mock Exams

The CKNE is the CNCF's new hands-on Kubernetes networking certification. Join Nico Vibert, part of the team that built it, for a practical workshop: work through real exam-style tasks in a live lab and leave ready to certify.

/authors/christian-hernandez

Christian Hernandez, Technical Marketing Engineer

/authors/katie-meinders

Katie Meinders, Community Builder

/authors/marino-wijay

Marino Wijay, Solutions Engineer