↑ Resources / Regulatory · Strategic

AI Agents in Critical Infrastructure: How NIST, the EU AI Act and NIS-2 Line Up

Regulatory · Strategic · 6 min read ·

On 7 April 2026, NIST published a concept note for a new profile of its AI Risk Management Framework, focused on trustworthy AI in critical infrastructure. It is an early document, not a finished standard, and NIST is collecting feedback before it drafts the profile. But it already shows what the agency expects from AI systems that help run power grids, water supplies and hospitals.

Operators in Europe will find much of it familiar. The EU AI Act and NIS-2 approach the same systems from different directions and arrive at similar expectations. This piece walks through what each framework asks for, where they overlap, and what that means for teams putting AI agents into critical environments.

What NIST is building

The profile applies the existing AI RMF, with its four functions of Govern, Map, Measure and Manage, to the conditions of critical infrastructure. The concept note covers AI across information technology, operational technology and industrial control systems, as well as the interfaces between them.

The note's examples make the expectations concrete. They include AI agents for autonomous cybersecurity incident response with tested and verified guardrails, diagnostic assistants that give traceable and auditable recommendations, autonomous robots and vehicles with deterministic fail-safe controllers, and optimization systems that degrade gracefully under adverse conditions while alerting human supervisors.

The common thread is that an operator has to know what an AI system did, why it did it, and what happens when it fails. In critical infrastructure, a failure is not a poor user experience. It can mean a power outage or a disrupted water supply.

What the EU AI Act requires

The EU AI Act classifies AI systems used as safety components in the management and operation of critical infrastructure, such as electricity, water, gas, heating and road traffic, as high-risk under Annex III.

High-risk systems must record events automatically over their lifetime (Article 12) and must be designed so that people can oversee them effectively (Article 14). Deployers have to keep the automatically generated logs for at least six months (Article 26). Under the Digital Omnibus adopted in 2026, these obligations apply to Annex III systems from 2 December 2027.

What NIS-2 requires

NIS-2 covers essential and important entities, including operators in energy, transport, water and health, through each member state's transposition law. Article 21 requires risk-management measures that include incident handling and supply chain security. Article 23 sets reporting deadlines: an early warning within 24 hours of becoming aware of a significant incident, an incident notification within 72 hours, and a final report within a month.

NIS-2 does not mention AI agents. But if an agent is involved in a significant incident, the operator still has to explain what happened within those deadlines, and that depends on having a usable record of what the agent did.

Where the frameworks overlap

The three frameworks were written independently, by different bodies, on different timelines. They use different language, and their duties are not identical. In practice, though, they ask operators for a similar set of capabilities:

  • A record of what each AI system did, kept long enough to investigate an incident.
  • The ability for people to understand, oversee and intervene in what the system does.
  • Predictable behavior when something goes wrong, including safe failure and alerts to humans.

A team that builds these capabilities once can draw on the same evidence for more than one framework.

Why critical infrastructure makes this harder

Critical infrastructure is rarely one system. A grid operator's agents may read from sensors, write to control systems, talk to trading platforms and log to enterprise systems, often at the same time and at machine speed.

Each of those layers has its own access controls, logging format and rules. The IT layer may fall under NIS-2. The OT layer may also fall under sector rules, such as Germany's KRITIS regulations. Industrial control components are often subject to safety standards written long before today's AI systems. None of these layers was designed to produce one consistent record of what an AI agent did across all of them.

That is why we think the record of agent actions belongs in a layer outside the individual systems, between the agents and the systems they act on. A record kept independently of the systems it describes is easier to rely on when an incident is investigated.

What this means for vendors

The concept note is written for operators, but its expectations will travel up the supply chain. Operators will ask the vendors of AI systems to show, not just state, that those systems meet them. A vendor's own dashboard is a weak answer on its own, because it is a record kept by the party being evaluated.

What to do now

The NIST profile is still being developed. The EU AI Act's high-risk obligations for these systems apply from 2 December 2027. NIS-2 already applies. That leaves time, but not much, to answer three questions for every agent in a critical environment: what it is allowed to do, whether you could reconstruct one of its past actions from start to finish, and where that record lives.

Operators who answer those questions now will be building the capability on their own schedule. Those who wait for a deadline will be building it under pressure.

KYDE

Kyde Control sits between your agents and the systems they act on. It scores every call against the agent's own baseline, holds actions that don't fit until a person releases them, and keeps a signed record of what each agent did. That record supports the agent layer of your EU AI Act and NIS-2 evidence. It does not discharge those duties on its own.