An SP 800-53 baseline is written for a generic system, and most real systems are not generic. High-performance computing clusters, industrial control systems, AI agents and systems holding classified or personal data all need controls the baselines leave out, and some baseline controls that do not fit at all. Overlays are NIST's answer: a reusable, documented set of tailoring decisions for a community, technology or operating environment, applied on top of a baseline. This guide explains what an overlay contains, how to read and combine published overlays such as SP 800-234 for HPC, where NIST's AI overlay project stands, and how to build your own overlay for a class of systems you run many times. It is for security architects, authorising officials, assessors and GRC teams working with the NIST Risk Management Framework, FedRAMP or an SP 800-53 based control set. Status was checked on csrc.nist.gov on 27 October 2026.
Key takeaways
- An overlay is tailoring you can reuse: it adds, removes, refines and parameterises controls for a defined type of system, so every system of that type does not repeat the same analysis.
- Overlays sit between baseline and system: select the SP 800-53B baseline, apply one or more overlays, then do system-specific tailoring under PL-11 Baseline Tailoring.
- SP 800-234 is the newest worked example: the final HPC Security Overlay (May 2026) starts from the 287-item moderate baseline and gives HPC guidance on 60 controls across four zones.
- The AI overlays are still in development: NIST's project, now named Control Overlays for Super Intelligence Systems (COSIS), has a concept paper (August 2025) and an annotated outline (January 2026), with single-agent and multi-agent use cases planned.
- Your own overlay is worth building when you operate many similar systems (Kubernetes platforms, SaaS tenants, agent platforms): it makes authorisation faster and assessments consistent.
What an overlay is and where it sits
SP 800-53B "Control Baselines for Information Systems and Organizations" provides low, moderate and high security baselines and a privacy baseline, plus tailoring guidance and "guidance on the development of overlays to facilitate control baseline customization". The NIST concept paper for AI overlays describes the same idea: overlays let organisations customise the controls or control baselines for a specific technology, system or mission space. The live NIST SP 800-53 Guide covers baselines and the basic tailoring actions; this page goes into the overlay itself.
| Layer | Who writes it | Scope | Example |
|---|---|---|---|
| Control catalogue | NIST | All controls and enhancements | SP 800-53 Rev. 5 (Release 5.2.0 issued 27 August 2025) |
| Baseline | NIST | Impact level (low, moderate, high) or privacy | SP 800-53B moderate baseline |
| Overlay | NIST, government bodies, sector groups, or your organisation | A community, technology or environment | SP 800-234 HPC Security Overlay |
| System tailoring | System owner, approved by the authorising official | One system | The system security plan and its tailoring decisions |
Counting from NIST's machine-readable OSCAL baseline profiles (usnistgov/oscal-content, read 27 October 2026), the baselines hold 149 controls and enhancements at low, 287 at moderate, 370 at high, and 96 in the privacy baseline. An overlay changes that list for its context, so the numbers above are the starting point an overlay author works from.
What an overlay contains
A usable overlay is more than a list of control IDs. When you read or write one, check that it covers these elements:
- Identification: name, version, owner, date and the catalogue and baseline release it was built against. An overlay built on Rev. 4 cannot be applied mechanically to Rev. 5.
- Characteristics and applicability: what kind of system or environment it is for, and the questions that decide whether a system qualifies. SP 800-234, for example, assumes the four-zone HPC reference architecture from SP 800-223.
- Base: which baseline it modifies. SP 800-234 builds on the moderate baseline and encourages organisations to adapt it, for example for multi-tenancy or a high baseline.
- Overlay summary: a table of every affected control and what the overlay does to it.
- Detailed control specifications: the supplemental guidance, discussion, parameter values and justification for each change.
- Tailoring considerations: what system owners may still change, and how this overlay interacts with others.
The types of change an overlay can make
| Change | What it means | Example |
|---|---|---|
| Add a control | Bring in a control or enhancement the baseline does not include | SP 800-234 adds AC-10 Concurrent Session Control, outside the moderate baseline, because HPC users run many concurrent sessions |
| Supplemental guidance | Explain how a baseline control applies in this context | SP 800-234 guidance on AU-2 Event Logging: parallel jobs can duplicate log entries, so check for duplicates and prioritise logging by zone |
| Discussion or compensating control | Accept that a control cannot be met as written and describe an alternative | SP 800-234 notes that encrypting internal HPC traffic under SC-8 may not be practical and points to compensating controls such as private, non-routable networks |
| Parameter values | Set organisation-defined parameters for the whole community | The AI concept paper notes that parameter values for assignment and selection operations can be set in an overlay |
| Remove or scope out | Mark a control as not applicable for the context, with justification | A control for a capability the system type never has |
Published overlays to know
| Overlay | Owner and status (27 Oct 2026) | Use it for |
|---|---|---|
| SP 800-234 High-Performance Computing Security Overlay | NIST; final, May 2026 (draft May 2025) | HPC clusters and research computing |
| Control Overlays for Super Intelligence Systems (COSIS), formerly COSAiS | NIST; concept paper 14 August 2025, annotated outline January 2026; no overlay draft listed on the project page | Generative AI assistants, predictive AI, single-agent and multi-agent systems, AI developers |
| Operational technology overlay | NIST, published within SP 800-82 Rev. 3 (September 2023); Rev. 4 initial public draft open for comment until 30 November 2026 | Industrial control and OT systems. See the NIST SP 800-82 OT Security Guide. |
| CNSSI 1253 and its overlays (for example privacy and classified information) | Committee on National Security Systems | US national security systems |
| Control Overlay Repository | NIST; government-wide, public and NIST categories; page updated 24 July 2026 | Finding overlays from other agencies and sectors before writing your own |
FedRAMP baselines are a related but separate idea: they are tailored baselines for cloud services rather than overlays in the SP 800-53B sense, although they are built and read the same way. See the FedRAMP Guide.
SP 800-234 as a model
SP 800-234 is worth reading even if you never run HPC, because it shows overlay technique cleanly. It defines five tailoring levels: All Zones plus each of the Access, Management, Computing and Data Storage zones. Its summary table marks each control "G" (supplemental guidance), "D" (discussion added), "--" (no change) or "N" (outside the moderate baseline but added). Its identity guidance is specific: IA-2(1) and IA-2(2) multi-factor authentication for system access, with node-to-node connections allowed without MFA under the same identity used for single sign-on; IA-11 Re-authentication often tailored because forcing re-authentication breaks long-running jobs; and AC-2 Account Management guidance that an account's role should dictate its access, with the Management Zone reachable only by administrators.
The AI overlays and the agent angle
NIST's project page states that on 20 October 2026 the project was renamed from SP 800-53 Control Overlays for Securing AI Systems (COSAiS) to Control Overlays for Super Intelligence Systems (COSIS), following a 29 September 2026 executive order directing agencies to use "Super Intelligence" in place of "Artificial Intelligence" in official materials; documents already issued keep their original wording. The substance has not changed: overlays built from SP 800-53 controls, drawing on SP 800-218A, draft NIST AI 800-1 and NIST AI 100-2 E2025, for five use cases:
- Adapting and using generative AI: assistant or LLM.
- Using and fine-tuning predictive AI (the January 2026 annotated outline covers this one).
- Using AI agent systems: single agent. The concept paper's examples are an enterprise copilot connected to email, files, calendar and internal systems that uses MCP to reach data sources, and a coding assistant that edits files, commits pull requests, runs tests and helps with deployment.
- Using AI agent systems: multi-agent, for example an expense workflow where one agent extracts claim data, another validates it against policy using retrieval, and another routes approved claims for payment. The concept paper calls this the least mature use case and mentions MCP and A2A for agent-to-agent communication.
- Security controls for AI developers.
The concept paper is explicit that the overlays assume an enterprise control set is already in place, naming access control for datasets and services, account management, and identification and authentication. That matters for identity teams: the AI overlays will not fix weak agent identity on their own. Until they are published, organisations deploying agents can build an interim overlay. Candidate controls to supplement and parameterise:
| Agent risk | SP 800-53 control to tailor | Overlay content to write |
|---|---|---|
| Agent acting with a user's full rights | AC-6 Least Privilege | Agents get their own scoped identity and task-limited permissions, never the invoking user's standing rights |
| Agents and tools authenticating to each other | IA-9 Service Identification and Authentication | Workload credentials for agents and MCP servers, short-lived tokens, no shared API keys |
| Agent accounts created ad hoc | AC-2 Account Management | Owner, purpose, expiry and review cycle for every agent identity |
| Actions not attributable | AU-2 Event Logging, AU-12 Audit Record Generation | Log the agent identity, the delegating human, the tool called and the data touched |
| Agent reaching systems it should not | AC-4 Information Flow Enforcement | Allowed tool and data destinations per agent class |
| Unexpected agent behaviour | SI-4 System Monitoring | Baselines of normal tool use and alerts on deviation |
See the Agent Identity and Access Guide and the NIST Cyber AI Profile Guide.
Applying several overlays to one system
A system can qualify for more than one overlay: an AI training cluster may be both HPC and AI, a plant historian both OT and privacy-relevant. Rules that keep this manageable:
- Order: select the baseline (PL-10 Baseline Selection), apply overlays, then do system-specific tailoring (PL-11). Record the order in the system security plan.
- Additions accumulate: if any applicable overlay adds a control, the system includes it.
- Conflicts go to the authorising official: where one overlay relaxes a control another strengthens, decide in the plan, with the risk reasoning. Do not let tools silently pick one.
- Parameter values: where overlays set different values, take the stricter unless the risk owner records why not.
- Versioning: pin each overlay version in the plan, and review when the overlay or the catalogue release changes.
Building your own overlay
Write an overlay when you run many systems of one type and keep making the same tailoring decisions. Typical candidates: a container platform, a standard SaaS tenant build, an agent platform, a data science environment, a branch office network.
- Define applicability. Describe the system type and the yes or no questions that decide whether a system qualifies. Include the reference architecture, as SP 800-234 does with its zones.
- Pick the base baseline. Usually moderate. Record the catalogue release (for example SP 800-53 Release 5.2.0).
- Threat-model the system type. Use SP 800-30 style analysis to find what the baseline misses and what it over-specifies. See the NIST SP 800-30 Risk Assessment Guide.
- Walk every baseline control. Mark it unchanged, guidance, discussion or compensating control, parameter set, or not applicable, with a reason for every change.
- Add what is missing. Bring in controls and enhancements outside the baseline, for example AC-2(12) or SR-3 for a platform with heavy third-party components.
- Write it in OSCAL. A profile that imports the baseline and modifies it makes the overlay machine-readable and lets system plans inherit it. See the OSCAL Guide.
- Pilot on two systems. Build the security plans, run a 800-53A assessment on each and note where assessors disagree with the overlay. See the NIST SP 800-53A Guide.
- Approve and publish internally. The risk executive or authorising officials approve the overlay; publish it with a version and owner.
- Maintain and roll back. Review at each catalogue release and after incidents. Keep the previous version available so systems can stay on it until re-assessed.
Small versus large organisations
- Small teams rarely need their own overlay. Use NIST overlays that fit, and keep system-level tailoring in one well-written plan.
- Large or federal estates gain the most: one approved overlay can remove weeks of repeated debate per authorisation and make continuous monitoring consistent across systems. See the NIST Risk Management Framework (SP 800-37) Guide.
Measuring whether overlays work
- Time from system registration to authorisation for systems using an overlay versus those without.
- Number of system-specific tailoring changes per system on top of the overlay (falling means the overlay fits).
- Assessment findings per system against overlay controls, and how many trace back to unclear overlay guidance.
- Percentage of qualifying systems actually using the current overlay version.
- Age of the overlay against the current catalogue release.
Framework mapping
| Framework | Where overlays fit |
|---|---|
| NIST SP 800-53 Rev. 5 | PL-10 Baseline Selection, PL-11 Baseline Tailoring, PM-9 Risk Management Strategy; the overlay content itself lives in the tailored controls |
| NIST RMF (SP 800-37) | Select step: baseline, overlays and tailoring; Monitor step uses the same tailored set with CA-7 Continuous Monitoring |
| NIST CSF 2.0 | GV.PO-01 policy based on organisational context; ID.RA-05 using threats, vulnerabilities, likelihood and impact to prioritise risk response |
| ISO/IEC 27001:2022 | The nearest equivalent is the Statement of Applicability. See the ISO 27001 Statement of Applicability Template. |
Common failure patterns
- Overlay as a shortcut to fewer controls: overlays written mainly to remove controls, without compensating controls, are rejected by assessors and leave real gaps.
- Stale baseline release: an overlay written against Rev. 4 and copied into Rev. 5 plans references withdrawn controls and misses new ones such as the SR family.
- No applicability test: systems adopt an overlay because it is convenient, not because they match its architecture assumptions.
- Silent conflicts: two overlays set different values and the GRC tool keeps whichever was loaded last.
- Identity left as assumed: overlays for new technologies, including the AI concept paper, assume account management and authentication are already solid. If they are not, the overlay inherits the weakness. Machine and agent accounts are where this shows first.
- Overlay without an owner: nobody reviews it, and it drifts from both the catalogue and the systems using it.
Practitioner checklist
- Check the NIST Control Overlay Repository and SP 800-234, SP 800-82 and COSIS before writing your own overlay.
- Record, for every system, the baseline, the overlays applied with versions, and the order of application.
- Require an applicability test for each overlay and keep the answers in the security plan.
- Resolve conflicts between overlays in the plan, with the authorising official's decision recorded.
- Build internal overlays for repeated system types, with a reason for every control change.
- Write overlays as OSCAL profiles so plans and assessments can inherit them.
- For AI agents, supplement AC-2, AC-6, IA-9, AU-2 and AC-4 now rather than waiting for the COSIS drafts.
- Review each overlay at every SP 800-53 release and after significant incidents.
Standards and references
- NIST: SP 800-53B, Control Baselines for Information Systems and Organizations
- NIST: SP 800-53 Rev. 5, Security and Privacy Controls
- NIST: SP 800-234, High-Performance Computing Security Overlay
- NIST: Control Overlays for Super Intelligence Systems (formerly COSAiS) project
- NIST: SP 800-53 Control Overlays for Securing AI Systems Concept Paper
- NIST: Control Overlay Repository
- NIST: SP 800-82 Rev. 3, Guide to Operational Technology Security
- NIST: OSCAL content, SP 800-53 Rev. 5 catalogue and baseline profiles
Related NHI Mgmt Group resources: NIST SP 800-53 Guide · NIST Risk Management Framework (SP 800-37) Guide · OSCAL Guide · Agent Identity and Access Guide · FedRAMP Guide