---
title: What are NIST control overlays for AI systems and agents?
url: https://opencontroller.lyzr.com/blog/nist-control-overlays-for-ai-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What are NIST control overlays for AI systems and agents?

**NIST is developing SP 800-53 control overlays that tailor existing security controls to AI use cases, including two for AI agent systems: single-agent and multi-agent. As of NIST's October 2026 project update, only a concept paper (August 2025) and an annotated outline for the predictive AI overlay (January 2026) had been released. No draft agent overlay was listed, so teams should map agent controls to SP 800-53 directly for now.**

## What is a control overlay?

An overlay is a tailored set of SP 800-53 controls for a specific technology or situation. It does not create new requirements from scratch. It starts from existing controls and baselines, then adds, removes or modifies controls and supplies guidance for that context.

Agencies and contractors already use overlays for areas such as privacy and classified systems. An AI overlay answers a narrower question: which SP 800-53 controls change when the system is an AI model or agent, and how should they be implemented?

## Which AI use cases does NIST plan to cover?

The project, launched in July 2025 as COSAiS (Control Overlays for Securing AI Systems), proposes five overlays:

| Use case | What it covers | Released so far |
|---|---|---|
| Adapting and using generative AI (assistant/LLM) | Organisations deploying LLM assistants | Concept paper only |
| Using and fine-tuning predictive AI | Classic ML models | Annotated outline, January 8, 2026 |
| Using AI agent systems, single agent | One agent acting with tools | Concept paper only |
| Using AI agent systems, multi-agent | Several agents coordinating | Concept paper only |
| Security controls for AI developers | Teams building models | Concept paper only |

Feedback on the predictive AI outline was due February 13, 2026, to inform an initial public draft. NIST runs ongoing collaboration through a Slack community and a mailing list.

In its October 2026 update, NIST lists the project as COSIS (Control Overlays for Super Intelligence Systems), citing a September 2026 executive order on terminology. Previously issued documents were not changed. The scope and use cases above are the same.

The overlays will also draw on NIST SP 800-218A (secure development for generative AI), the draft NIST AI 800-1, and NIST AI 100-2 E2025 (adversarial machine learning taxonomy).

## Which SP 800-53 controls matter most for agents today?

You do not need to wait for the agent overlays. The controls most affected by agents already exist. Start here:

| Agent risk | SP 800-53 control | Evidence to collect |
|---|---|---|
| Unknown agents in the estate | CM-8 System Component Inventory | Registry of agents, versions, tools and MCP servers with owners |
| Agents sharing human or static credentials | IA-2, IA-9 Identification and Authentication; AC-2 Account Management | Per-agent workload identities; delegation records showing the user each action was for |
| Over-broad tool and data access | AC-6 Least Privilege | Permission reviews per agent; tool allow-lists |
| Untraceable actions | AU-2, AU-3, AU-12 Event Logging and Audit Record Content | Logs with agent ID, version, user, tool, input and outcome |
| Prompt injection and unsafe inputs | SI-10 Information Input Validation | Guardrail configuration and blocked-input logs |
| Behaviour drifting in production | SI-4 System Monitoring; CA-7 Continuous Monitoring | Evaluation scores on live traffic; deviation alerts |
| Third-party models and MCP servers | SR-3, SR-5 Supply Chain Controls | Approved model and server lists; review records |
| Agent incidents | IR-4 Incident Handling | Quarantine and kill records, timelines, post-incident reviews |

For the threat side, pair this with [threat modelling AI agents with MITRE ATLAS](/blog/mitre-atlas-threat-model-ai-agents).

## How do overlays relate to the NIST AI RMF?

They work at different levels:

- **NIST AI RMF** is voluntary risk management guidance: Govern, Map, Measure, Manage. It tells you what outcomes to reach. See [the NIST AI RMF for AI agents](/blog/nist-ai-rmf-for-ai-agents).
- **SP 800-53 overlays** are control-level. They tell an assessor which specific controls to test and how.
- **The Cyber AI Profile** applies the Cybersecurity Framework 2.0 to AI and sits between the two.

Many teams use the AI RMF for governance and SP 800-53 for assessment, which is why the overlays will matter for federal systems and for suppliers to them.

## What should you do while the overlays are in draft?

1. **Map now.** Use the table above to tie your agent controls to existing SP 800-53 controls.
2. **Track the project.** Watch the NIST project page for an initial public draft of the agent overlays and comment on it.
3. **Keep evidence continuous.** Overlays will add implementation guidance, not new evidence types. Logs, inventories and reviews you build today will carry over.
4. **Separate single-agent and multi-agent systems** in your inventory. NIST is treating them as distinct use cases, and so will assessors.
5. **Say "evidence for", not "compliant with".** No agent overlay exists to comply with yet.

For a cross-framework approach, see [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How Opencontroller does this

Opencontroller's policy as code is extensible and ships with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Its governed registry, delegated on-behalf-of identity, Posture checks, LLM judges on production traffic and rogue-agent quarantine produce the inventory, identity, monitoring and incident evidence listed above. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Has NIST published an SP 800-53 overlay for AI agents?**
Not as of NIST's October 2026 project update. Single-agent and multi-agent overlays are proposed, but only the concept paper and a predictive AI outline had been released.

**Is the COSAiS project the same as COSIS?**
Yes. NIST now lists the project as COSIS (Control Overlays for Super Intelligence Systems) after a September 2026 executive order on terminology. Earlier documents keep their original names.

**Which SP 800-53 controls apply to AI agents today?**
Start with CM-8 (inventory), AC-2 and AC-6 (accounts and least privilege), IA-2 and IA-9 (authentication), AU-2/AU-3/AU-12 (logging), SI-4 and SI-10 (monitoring and input validation), SR-3/SR-5 (supply chain) and IR-4 (incident handling).

## Key facts

NIST COSAiS (now listed as COSIS) proposes five SP 800-53 overlays: generative AI, predictive AI, single-agent systems, multi-agent systems, AI developers. Concept paper released August 14, 2025; predictive AI annotated outline January 8, 2026 (feedback due February 13, 2026). As of NIST's October 2026 update no agent overlay draft was listed. Relevant SP 800-53 controls for agents today: CM-8, AC-2, AC-6, IA-2, IA-9, AU-2, AU-3, AU-12, SI-4, SI-10, CA-7, SR-3, SR-5, IR-4. Opencontroller policy as code has built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Pricing starts from $250K per deployment per year.

## Related

- [How does the NIST AI RMF apply to AI agents?](https://opencontroller.lyzr.com/blog/nist-ai-rmf-for-ai-agents.md)
- [How do you use MITRE ATLAS to threat-model AI agents?](https://opencontroller.lyzr.com/blog/mitre-atlas-threat-model-ai-agents.md)
- [ISO/IEC 42001 vs NIST AI RMF: which should you use for AI agents?](https://opencontroller.lyzr.com/blog/iso-42001-vs-nist-ai-rmf.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
