HPC AI Workloads Need Runtime Security. The Architecture Already Exists.

This report describes a series of high-impact AI-focused supply-chain attacks (compromised PyPI and npm packages and a trojanized signed vendor binary) and argues that existing NIST HPC security guidance is inadequate for AI workloads. It documents runtime threats — training-data poisoning, GPU side-channel exfiltration, and inference-time manipulation — and presents SentinelOne’s detection examples, architecture recommendations (dedicated security compute, eBPF behavioral telemetry, continuous model integrity verification), and a call for SR-family extensions such as an AI bill of materials.