Authors: Saber Feki (King Abdullah University of Science and Technology (KAUST)), Yang Guo (NIST), Stephen Green (King Abdullah University of Science and Technology (KAUST)), Kevin DeVilbiss (Alvarez and Marsal)
Abstract: HPC systems were built for trusted users. Today they run a variety of HPC and AI workloads, serve GPU-as-a-service customers they've never met, and face rising regulatory attention. NIST is finalizing SP 800-234, the first HPC-specific security overlay, and advanced compute is one of the few genuinely bipartisan topics in Washington. The rules are coming. The question is whether operators shape them or inherit them. This BoF brings standards authors, security architects, supercomputing operators, and implementers into a candid, practitioner-level exchange with the audience on what works, what fails, and how the community secures shared compute without slowing the science.
Long Description: For decades, HPC security rested on a single assumption: you knew who was on your machine. That assumption is gone. Systems built for trusted research communities now support multi-tenant HPC and AI workloads, GPU-as-a-service customers, and international collaborations, and they face rising expectations around identity, access, monitoring, usage and trust. At the same time, governments are actively moving toward legislation and standards for advanced compute infrastructure. The specifics are still forming, but the direction is clear, and this is one of the few technology topics drawing sustained bipartisan interest in Washington. The architecture choices operators make now will determine who is ready and who is retrofitting under pressure.
Most of these conversations happen in isolation. Policymakers debate objectives. Technology providers pitch capabilities. Investors size the build-out. Operators are left to translate all of it into systems that actually run jobs and keep researchers productive. This BoF flips that dynamic and puts the people who own and operate the infrastructure at the center.
The topic has already demonstrated strong pull across the ecosystem. In June 2026, the organizers convened more than 100 participants in Washington, DC to discuss the future of secure and scalable AI infrastructure. The room spanned technology providers, infrastructure operators, systems integrators, investors, security and legal practitioners, policymakers, and government officials, and included a sitting Member of Congress alongside senior congressional staff. One theme surfaced repeatedly: no matter where policy originates or which technology wins, the hard implementation questions land on the operators of shared compute.
This session puts those questions directly to the HPC community:
• The old HPC security model and what replaces it, including workload isolation on shared accelerators, side-channel risk, and memory residue across jobs.
• What the first HPC security overlay (NIST SP 800-234) gets right, and where operators still need to go further on multi-tenant isolation and continuous compliance.
• Security controls researchers will actually use, since most incidents are legitimate users routing around controls that slowed them down.
• The unresolved collision between open research and use and access restrictions on shared systems.
• Continuous verification and the architecture decisions operators should be making now, ahead of final rules.
• Access governance and know-your-customer for compute when resellers, brokers, and API aggregators sit between the owner and the end user.
The session opens with short perspectives from leaders across standards development, security architecture, large-scale supercomputing operations, and implementation, then moves to open discussion, experience sharing, and identification of the most pressing challenges facing operators over the next few years.
This is a new BoF for SC26. It builds on a webinar and an in-person working session held earlier in 2026, and it is designed to give the HPC community a direct seat in shaping security and governance practices for shared HPC and AI infrastructure before those decisions are made elsewhere.