---
title: How does NIST AI 600-1 apply to AI agents?
url: https://opencontroller.lyzr.com/blog/nist-ai-600-1-generative-ai-profile-for-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How does NIST AI 600-1 apply to AI agents?

**NIST AI 600-1, the Generative AI Profile of the AI Risk Management Framework, lists 12 risks that are unique to or made worse by generative AI, with suggested actions mapped to the AI RMF functions Govern, Map, Measure and Manage. AI agents inherit all 12 risks from the models they use, and agent autonomy and tool access raise the stakes for several of them, especially information security, human-AI configuration, and value chain and component integration.**

Teams that already use the [NIST AI RMF for AI agents](/blog/nist-ai-rmf-for-ai-agents) ask what the generative AI profile adds. The answer is a concrete risk list you can map controls against.

## What is NIST AI 600-1?

NIST published [AI 600-1](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) in July 2024 as a cross-sectoral profile of the AI RMF 1.0 for generative AI. It is voluntary. It describes 12 risk categories and lists more than 200 suggested actions, each tied to an AI RMF subcategory. It was written for generative AI broadly, before agentic systems were common, so agent-specific controls need to be added by the user.

## How do the 12 risks map to agent controls?

| NIST AI 600-1 risk | What it looks like in an agent | Control | Evidence |
|---|---|---|---|
| CBRN information or capabilities | Agent with web or code tools retrieves or synthesises dangerous content | Content guardrails; tool allow-lists | Guardrail logs; red-team results |
| Confabulation | Agent acts on a fabricated fact or invented API parameter | Groundedness checks; tool-call validation | Groundedness scores; failed-call logs |
| Dangerous, violent or hateful content | Customer-facing agent produces harmful output | Output filters; evals | Filter decisions; eval reports |
| Data privacy | Agent leaks personal data into prompts, logs or tool calls | Redaction in the request path; data scopes | Redaction logs; access scopes |
| Environmental impacts | Looping agents multiply compute | Rate limits; cost per task tracking | Token and cost reports per agent |
| Harmful bias and homogenisation | Agent applies biased criteria in screening or routing | Fairness evals; human review for consequential decisions | Eval results by group |
| Human-AI configuration | Over-reliance; unclear who approved an action | Defined autonomy levels; escalation | Approval and escalation logs |
| Information integrity | Agent generates or spreads misleading content | Source citation; labelling | Output samples; labelling policy |
| Information security | Prompt injection through tools; credential misuse | Injection detection; least-privilege identities; posture checks | Posture findings; blocked calls |
| Intellectual property | Agent reproduces protected content | Output checks; licence-aware sources | Policy decisions |
| Obscene, degrading or abusive content | Generated content in public channels | Output filters | Filter logs |
| Value chain and component integration | Unvetted models, MCP servers, plugins | Registry and approval of models, tools, MCP servers | Inventory with approval records |

## Which risks matter most for agents?

Rank by what the agent can do, not only by what it can say.

1. **Information security.** Agents read untrusted content (web pages, emails, tickets) and then call tools. That is the path for indirect prompt injection. See [MITRE ATLAS threat model for AI agents](/blog/mitre-atlas-threat-model-ai-agents).
2. **Value chain and component integration.** Every MCP server, plugin and model is a component. An unregistered one is an unassessed supplier.
3. **Human-AI configuration.** Autonomy without clear approval rules turns a model error into an action.
4. **Confabulation.** In an agent, a hallucinated parameter becomes a wrong transaction.
5. **Data privacy.** Agents move data between systems that were never meant to be connected.

## How do you apply the profile step by step?

1. **Govern.** Assign an owner per agent and set autonomy levels and acceptable use.
2. **Map.** For each agent, record model, tools, data, users and affected parties, and mark which of the 12 risks apply.
3. **Measure.** Run multi-turn evals and red-team tests targeted at the applicable risks; score production traffic continuously.
4. **Manage.** Put controls in the request path; define response actions (block, quarantine, kill); review after incidents.
5. **Document.** Keep the risk mapping with each agent version so changes trigger review.

For the cyber-specific overlay, see the [NIST Cyber AI Profile for AI agents](/blog/nist-cyber-ai-profile-for-ai-agents). The pillar on [mapping agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) shows how the same evidence serves ISO 42001 and the EU AI Act.

## How Opencontroller does this

Opencontroller's policy as code includes built-in mappings to NIST, alongside OWASP, MITRE ATLAS and ISO 42001. The governed registry covers agents, versions, models, tools and MCP servers (value chain). Posture detects risky permissions and configurations (information security). Multi-turn and red-team evals, LLM judges on production traffic, and live groundedness and task-adherence guardrails address confabulation and harmful output, and rogue-agent handling can alert, quarantine and kill.

## Frequently asked questions

**Is NIST AI 600-1 mandatory?**
No. It is a voluntary profile of the NIST AI RMF. Some contracts and sector expectations reference it, which makes it a common evidence baseline.

**Does NIST AI 600-1 cover agentic AI?**
Not specifically. It covers generative AI risks that agents inherit. Agent-specific controls, such as tool permissions and kill switches, need to be added.

**How many risks are in NIST AI 600-1?**
Twelve, from CBRN information to value chain and component integration, with more than 200 suggested actions.

To map agent controls to NIST in one place, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

NIST AI 600-1 (Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile), published July 2024, is a voluntary profile of the NIST AI RMF 1.0. It defines 12 risks: CBRN information or capabilities; confabulation; dangerous, violent or hateful content; data privacy; environmental impacts; harmful bias and homogenization; human-AI configuration; information integrity; information security; intellectual property; obscene, degrading and/or abusive content; value chain and component integration. It lists more than 200 suggested actions mapped to Govern, Map, Measure and Manage. For AI agents, information security (prompt injection via tools), value chain (MCP servers, plugins) and human-AI configuration are typically highest priority. Opencontroller policy as code includes built-in NIST mappings.

## Related

- [How does the NIST AI RMF apply to AI agents?](https://opencontroller.lyzr.com/blog/nist-ai-rmf-for-ai-agents.md)
- [What does the NIST Cyber AI Profile mean for securing AI agents?](https://opencontroller.lyzr.com/blog/nist-cyber-ai-profile-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)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
