EAGER: Unmasking HPC Abuse: AI Graph Inference from Scheduling Metadata

This research security project aims to protect AI workloads, supercomputing cyberinfrastructure, and embargoed research data from sophisticated threats. Recent attacks have shown that these threats can be subtle, often disguised as routine maintenance or accidental system failures, challenging security operators. Despite having subtle traces, these attacks can significantly disrupt research momentum, exfiltrate sensitive data, corrupt scientific findings, and ultimately undermine public confidence in the critical infrastructure driving the AI innovation engine. Traditionally, mitigating such evolving threats requires significant efforts to curate historical attack traces and discover out-of-distribution lateral movements. By leveraging high-performance computing for accelerated analytics, this project will develop a self-securing AI infrastructure protected by AI agents. Additionally, this project will rigorously educate students, research administrators, and scientists about insidious cyber-threats, allowing them to mitigate risks associated with future AI workloads powering research innovations. The study focuses on uncovering improper uses of resources in supercomputing cyberinfrastructure, leveraging the National Center for Supercomputing Applications (NCSA) as the main vantage point. The technical approach involves deploying a federation of AI agents to process unstructured logs, including high-speed interconnects, login hosts, and GPU nodes from a data lake such as AICyberLake. This approach provides statistical insights into utilization, job completion, energy consumption, temperatures, and graphs of scientific workflow metadata using job schedulers, e.g., SLURM or PBS. The aim is to pinpoint uncertainties and uncover new research security violations exploiting emerging technologies such as AI-driven malware and quantum-resistant cryptography communications. A product of the study will be a standardize knowledge base of stealthy attack/abuse techniques on graphics processing unit (GPU)-accelerated systems, working with the National Institute of Standards and Technology (NIST), to provide a blueprint of such activities and corresponding mitigations. Successful implementation will yield novel graph-based AI agents inference methods and provide concrete attack case studies to inform future research. The project will also contribute to course materials on research security that will be broadly applicable to other research computing centers, ultimately unleashing a trustworthy AI innovation engine. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.