Google Cloud Unveils New Blueprint for Securing AI Workloads on GKE

Google Cloud published a security blueprint for securing AI workloads on Google Kubernetes Engine (GKE), authored by Glen Messenger and Shannon Kularathna. The document targets CISOs and platform engineering teams, proposing a three-layer security approach covering infrastructure, model integrity, and application security. At the infrastructure layer, it recommends Confidential GKE Nodes, which extend hardware-level memory encryption to NVIDIA H100 GPUs and TPUs. For model security, the blueprint introduces k8s-aibom, an open-source Kubernetes controller that automatically generates an AI bill of materials, addressing assets like datasets and ML frameworks beyond traditional SBOMs. At the application layer, Model Armor inspects prompts and responses for injection attacks, sensitive data leaks, and harmful content, while GKE Sandbox based on gVisor isolates AI agents that execute generated code or call untrusted tools. Google suggests a phased deployment: Deploy (basic controls), Operate (production hardening), and Govern (organizational guardrails and automated incident response). The article notes that AWS has a similar approach with its AWS AI Security Framework and AI on EKS project, and Microsoft focuses on AI agent identity via Microsoft Entra Agent ID and the PyRIT red-teaming tool.