AI infrastructure concentrates valuable assets in one place: proprietary model weights, sensitive training data and expensive accelerators. It is also built from components that were designed for trusted research environments, such as distributed compute frameworks, experiment trackers and notebook servers, several of which have had serious vulnerabilities or been found exposed to the internet (MITRE ATLAS added a technique for misconfigured or publicly exposed AI services in its September 2026 release). This course applies infrastructure security practice to that environment, drawing on Wirewalk's HPC and AI infrastructure work.
You will assess GPU multi-tenancy options (MIG, time-slicing, whole-node allocation) and their isolation limits, harden Slurm and Kubernetes for AI workloads, control access on InfiniBand and RoCE fabrics and parallel file systems, and manage node images, drivers and firmware. You will harden inference servers such as vLLM and Triton, put gateways, authentication and quotas in front of endpoints, and protect weights at rest, in transit and in use. The MLOps module secures notebooks, experiment trackers, registries, CI runners and feature stores. You will evaluate GPU confidential computing with attestation and air-gapped or sovereign deployments, design telemetry, map controls to NIST SP 800-53, the draft NIST Cyber AI Profile (IR 8596) and CIS Benchmarks, and finish with a capstone review of a hypothetical 64-node training cluster. The Wirewalk Institute certificate attests to completion and assessment of this course; it is not endorsed by OWASP, MITRE, NIST or any vendor named here.