AI agent workloads fundamentally change the Kubernetes security landscape by introducing unpredictable traffic, ephemeral GPU-intensive processes, and dynamic tool discovery. Azure Kubernetes Service addresses this across four control planes: network isolation via network-isolated clusters and private registries to prevent unauthorized egress; policy-as-code using OPA and Kyverno for granular agent-to-agent authorization including emerging standards like transaction tokens and AAuth; image scanning and provenance validation with curated registries, continuous vulnerability scanning, and lightweight sandboxing via Hyperlight (1-2ms startup); and runtime anomaly detection using eBPF-based monitoring tuned for ephemeral agentic behavior. Secure GPU multi-tenancy is highlighted as an underappreciated challenge, with confidential GPUs still in preview. The article argues that platform-enforced security defaults benefit both agentic and traditional workloads sharing the same cluster.
