Enterprise security baselines break high-performance computing: idle session timeouts kill week-long jobs, endpoint scanners slow parallel file systems, and per-node logging floods the SIEM. NIST SP 800-234, the High-Performance Computing (HPC) Security Overlay, published in final form on 4 May 2026, tailors the SP 800-53B moderate baseline so HPC operators can stay secure without losing the performance their users need. It builds on the four-zone reference architecture of SP 800-223 and adds HPC guidance to 60 controls. This guide explains the zone model, the tailoring decisions that matter most, identity and non-human identity gaps the overlay leaves to you, an implementation plan and how to evidence it. It is for HPC system administrators, security managers and assessors at research centres, national labs, universities and organisations building GPU clusters for AI training.
Key takeaways
- It is a tailoring of moderate, not a new baseline: the overlay starts from the SP 800-53B moderate baseline, tailors 60 controls with HPC supplemental guidance and adds one control, AC-10 Concurrent Session Control.
- Everything is organised by zone: guidance is given for All Zones and separately for the Access, Management, High-Performance Computing and Data Storage zones defined in SP 800-223.
- Relaxations come with compensations: longer session limits, lighter internal encryption and trimmed logging are permitted only where the zone model and compensating controls carry the risk.
- The Management Zone gets the strictest treatment: admin-only access through a defined gateway or bastion, normal re-authentication and the highest logging priority.
- Machine identities are mostly left to you: the overlay does not address service accounts, SSH keys, Kerberos or scheduler credentials, which are where real HPC and AI cluster compromises start.
What it is and who it applies to
SP 800-234 is a security overlay in the sense of SP 800-53B: a fully specified set of control tailoring for a community of interest. NIST released the initial public draft on 1 May 2025 and the final on 4 May 2026, developed by its HPC Security Working Group with authors from national laboratories, the DoD HPC Modernization Program, NASA, MIT Lincoln Laboratory and universities. It is a companion to SP 800-223 (final, 9 February 2024), which provides the HPC reference architecture, threat analysis and security posture.
The overlay is mandatory only where a federal agency or contract requires SP 800-53 controls on an HPC system, for example a moderate impact federal system or an HPC service assessed for authorisation. Everyone else uses it as an authoritative, peer reviewed statement of which enterprise controls need adjusting for HPC and why. NIST names IT security managers, compliance officers, HPC system administrators and agency programme managers as the audience. It also applies, by analogy, to private GPU clusters used for large model training, which share the same architecture of login nodes, batch schedulers, high speed interconnects and parallel storage. See the NIST Control Overlays Guide for how overlays work in general.
Structure and key requirements
Section 1 introduces the zones and the five tailoring levels. Section 2 is the summary table marking every tailored control. Section 3 contains the tailored specifications in 14 topic groups (role-based access control, logging, user sessions, contingency planning, network connections, identification and authentication, emergency handling, user-developed software, performance and scalability, controls inapplicable to HPC, shared GPUs and accelerators, training and tailoring, management and maintenance, and access to HPC).
The four zones
| Zone | Who reaches it | Overlay emphasis |
|---|---|---|
| Access Zone (login and portal nodes) | All authorised users and administrators | Login sessions, MFA at entry, DoS detection (SC-5), all end-user traffic routes through it (AC-4) |
| Management Zone | System administrators only, through an approved gateway or bastion (AC-17(4)) | Highest logging priority, normal re-authentication (IA-11), separate networks or VLANs for login and API access |
| High-Performance Computing Zone (compute nodes) | Administrators, and users only on nodes the batch scheduler assigns | Long running jobs: no inactivity logout while jobs run (AC-2(5), AC-12), GPU data cleared between users (SC-4) |
| Data Storage Zone | Administrators only; users reach data through service interfaces such as a mounted POSIX file system | Admin re-authentication, storage-specific protection at rest (SC-28), selective backup (CP-9) |
Tailoring decisions that matter most
| Control | Enterprise default | HPC overlay guidance (summarised) |
|---|---|---|
| AC-2(5) Inactivity Logout, AC-12 Session Termination | Short idle timeouts | Thresholds longer than enterprise norms; never terminate sessions with running compute jobs |
| AC-10 Concurrent Session Control (added) | Not in moderate | Limits may vary by zone, set by risk assessment |
| AC-3 Access Enforcement | Role based | Access to one zone does not carry over to another; separate administrator duties |
| AC-6(5) Privileged Accounts | Restrict privileged accounts | Root-privileged admins should not run user-type jobs with elevated privilege in the compute zone; maintenance jobs may go through the scheduler |
| IA-2(1), IA-2(2) MFA | MFA for all access | MFA required for system access; internal non-MFA hops allowed only with the same identity and privilege; any identity or privilege change needs MFA |
| IA-11 Re-authentication | Periodic | Often effectively infinite for compute and access sessions; normal in the Management Zone; required for storage admins |
| AU-2, AU-4, AU-5 Logging | Log all defined events per host | De-duplicate events from parallel jobs, prioritise Management, then Access, then Storage, then Compute; central and preferably immutable logging; early alert on logging failure |
| AC-6(9) Log Use of Privileged Functions | Log privileged functions | Cutting privileged function logging carries more risk than trimming other events |
| SC-8 Transmission Confidentiality | Encrypt internal traffic | Internal encryption may be impractical on interconnects; compensate with private networks |
| RA-5, SI-3, CM-8(3) Scanning | Agent based scanning on every host | Limit overhead; scan one node or image for identical nodes; avoid repeated scans of shared file systems; prefer out-of-band or idle-time checks and real-time process scanning |
| SI-7 Integrity | Integrity checks on all software | System software only, not user software; full integrity checks on large storage often infeasible |
| CP-9 System Backup | Back up everything | Full backup often prohibitive; intermediate data may be excluded, and users must know what is and is not backed up |
| CM-7, CM-11 User software | Allow lists, no user installs | User software runs unprivileged in user space, isolated from Management and Storage zones; allow-by-exception lists often impractical in Access and Compute zones |
| SC-15, SC-18 | Collaborative devices, mobile code | Generally not applicable to HPC |
A careful reading shows minor counting inconsistencies in the final text (the summary table markers total 61 including AC-10, against the stated 60). Record which interpretation you used in your system security plan.
Identity and NHI requirements
The overlay's identity guidance is concentrated in section 3.6 and is about people: MFA at the Access Zone, acceptance of PIV for entry without re-prompting internally (IA-2(12)), SSO within the HPC system only where appropriate (IA-1), and re-authentication rules per zone. It does not tailor IA-5 Authenticator Management and does not discuss service accounts, SSH keys, Kerberos or scheduler credentials. In practice those are the credentials that move laterally across thousands of identical nodes. The following is NHI Mgmt Group guidance that fills that gap while staying consistent with the overlay's MFA rule that identity or privilege changes must be strongly authenticated:
- SSH keys: passwordless intra-cluster SSH is often how the "same identity and privilege" hops in IA-2(1) are implemented. Generate per-user keys inside the cluster, never accept user-supplied private keys for cross-site access, and prefer short-lived SSH certificates issued after MFA at the Access Zone. See the SSH Key Management Guide.
- Kerberos and keytabs: many centres use Kerberos for parallel file system and node authentication. Treat host and service keytabs as secrets with owners, rotation and inventory; see the Linux Service Accounts and Keytab Management Guide.
- Scheduler and munge-style shared secrets: the batch scheduler decides which nodes a user may touch (AC-2 in the compute zone), so its authentication key is effectively a tier 0 secret. Restrict it to Management Zone hosts and monitor its use.
- Service and workflow accounts: data movers, monitoring agents, science gateways and workflow engines run under service accounts that bypass interactive MFA. Give each a named owner, the narrowest zone access and no interactive login. See the Service Account Security Guide.
- Cloud and AI API tokens in user jobs: AI training jobs pull data and models with cloud keys and model hub tokens stored in home directories or environment variables. The ShadowRay campaign showed attackers harvesting exactly these tokens, plus SSH keys, from AI clusters whose job submission API had no authentication. Prefer federated, short-lived credentials; see the AI Infrastructure Workload Identity Guide.
- Federated external users: where IA-2(12) allows alternate MFA for external collaborators, map federated identities to local accounts with expiry tied to project allocation; the Research Institution Identity Security Guide covers allocation-based lifecycle.
Implementation steps
- Categorise and confirm applicability: confirm the system is moderate impact (FIPS 199) or that a contract requires SP 800-53. If high impact, use the overlay's reasoning but start from the high baseline. See the NIST SP 800-53B Control Baselines and Tailoring Guide.
- Map your architecture to the four zones: list every node class (login, data transfer, compute, GPU, management, storage servers, scheduler, monitoring) and assign it to a zone. Ambiguous systems, such as science gateways, usually belong to the Access Zone.
- Enforce zone boundaries first: route all user connections through the Access Zone (AC-4), place management interfaces behind a bastion on separate networks (AC-17(4)), and prevent users logging in to storage servers.
- Apply identity controls: MFA at Access Zone entry and for all administrator access, step-up for any privilege change, separate admin accounts that do not run user jobs (AC-6(5)), and an inventory of SSH keys, keytabs and service accounts.
- Tune sessions and re-authentication per zone: configure job-aware session handling so running jobs survive disconnection (SC-10) while Management Zone sessions keep normal timeouts.
- Redesign logging: centralise logs (AU-4), de-duplicate parallel job events, keep full privileged and Management Zone logs, and alert on logging pipeline failure (AU-5), which matters for diskless nodes.
- Adapt scanning and integrity: scan golden node images rather than every node, run scans in idle windows, and scope SI-7 to system software.
- Handle shared accelerators: clear GPU memory and accelerator state between jobs from different users (SC-4), particularly for multi-tenant AI training.
- Write the contingency position: document what is backed up and what is not, and tell users (CP-1, CP-9). Consider peer institutions or cloud HPC as alternate sites (CP-7).
- Pilot and roll back safely: trial new session, logging and scanning settings on a test partition with representative long jobs before production, and keep previous scheduler and PAM configurations under version control (CM-2(2)) so a change that hurts throughput can be reverted.
Small centre or large facility
A university cluster with a few administrators can adopt the zone model with VLANs, a single bastion and a central syslog, and accept the overlay's relaxations wholesale. A national facility or a commercial AI training cluster with external tenants should go further: separate admin identity provider, hardware-backed admin credentials, network monitoring designed for high bandwidth flows (SI-4) and formal accelerator sanitisation. The trade-off is always performance against assurance; the overlay's value is that it documents which relaxations are considered acceptable.
Evidence and assessment
Assessors familiar with HPC are preferred (CA-2(1)), and the overlay treats the four-zone system as one component, with node-to-node compute connections as authorised internal connections (CA-9). Evidence that works well:
| Area | Evidence | Metric |
|---|---|---|
| Zone separation | Network diagrams per zone, firewall and VLAN rules, bastion configuration | Number of direct paths to Management or Storage zones that bypass the bastion (target zero) |
| Authentication | IdP and PAM configuration showing MFA at Access Zone and for admins | Percentage of interactive logins with MFA at entry; admin accounts without MFA (target zero) |
| Machine credentials | Inventory of SSH keys, keytabs, service accounts with owners | Percentage of service credentials with an owner and rotation date; age of oldest scheduler key |
| Logging | Central log configuration, retention settings, failure alerts | Hours of Management Zone log gaps per quarter; time to detect logging pipeline failure |
| Vulnerability management | Golden image scan reports, patch windows aligned to job drain | Median days to patch critical vulnerabilities on login nodes |
| Accelerators | Scheduler prolog and epilog scripts that reset GPU state | Percentage of multi-tenant GPU jobs followed by a verified reset |
Record each overlay tailoring decision in the system security plan as required by PL-11 Baseline Tailoring, with the compensating control named where a control is relaxed.
Mapping to other frameworks
| Overlay theme | NIST SP 800-53 Rev 5 | NIST CSF 2.0 | Related reading |
|---|---|---|---|
| Zone separation and bastion | AC-4, AC-17(4), CA-9 | PR.IR-01 | Privileged Access Workstations Guide |
| Role and zone based access | AC-2, AC-3, AC-6(5) | PR.AA-05 | NIST SP 800-53 Guide |
| MFA and re-authentication | IA-2(1), IA-2(2), IA-2(12), IA-11 | PR.AA | SP 800-63B for authenticator assurance |
| Logging and monitoring | AU-2, AU-4, AU-5, AC-6(9), SI-4 | DE.CM-01 | SP 800-92 log management |
| Shared accelerators | SC-4 | PR.DS | Confidential AI and GPU Enclaves Guide |
| Machine credentials (gap) | IA-4, IA-5 (not tailored by the overlay) | PR.AA | SSH Key Management Guide |
Common pitfalls
- Taking the relaxations without the compensations: dropping internal encryption (SC-8) is only reasonable when the interconnect is genuinely private and the zone boundaries are enforced.
- Infinite sessions on login nodes: the overlay protects running jobs, not idle interactive shells. Use job-aware tooling (SC-10) rather than disabling timeouts globally.
- Unauthenticated job and dashboard APIs: the ShadowRay attacks on Ray clusters show what happens when a job submission interface is reachable without authentication. Treat any job API as an Access Zone service behind MFA.
- Admins running user jobs as root: AC-6(5) guidance exists because a privileged job on a compute node can reach every mounted file system.
- Logging cut where it matters: trimming duplicate compute events is fine; trimming Management Zone or privileged function logs removes the evidence you need after a credential compromise.
- Shared SSH keys and stale keytabs: because the overlay is silent on machine credentials, they often go unreviewed for years and become the lateral movement path.
Practitioner checklist
- Download SP 800-234 and SP 800-223 and record the version and date used in the system security plan.
- Assign every node class and service to one of the four zones.
- Put Management Zone access behind a bastion with MFA and separate admin accounts.
- Require MFA at Access Zone entry and step-up for any identity or privilege change.
- Inventory SSH keys, Kerberos keytabs, scheduler keys and service accounts with named owners.
- Configure job-aware session handling and keep normal timeouts in the Management Zone.
- Centralise logs, prioritise Management and privileged events, and alert on logging failure.
- Scan golden images and schedule patching around job drain windows.
- Reset GPU and accelerator state between users in multi-tenant partitions.
- Document every tailoring decision and its compensating control under PL-11.
Standards and references
- NIST: SP 800-234 High-Performance Computing (HPC) Security Overlay
- NIST: SP 800-234 full text (PDF)
- NIST: SP 800-223 High-Performance Computing Security: Architecture, Threat Analysis, and Security Posture
- NIST: High-Performance Computing Security Working Group
- NIST: SP 800-53 Rev 5 Security and Privacy Controls
- NIST: SP 800-53B Control Baselines
Related NHI Mgmt Group resources: NIST Control Overlays Guide · NIST SP 800-53B Control Baselines and Tailoring Guide · AI Infrastructure Workload Identity Guide · Research Institution Identity Security Guide