Idaho National Lab Deploys Autonomous AI to Guard Power Grid, Rules Stay Secret

Idaho National Laboratory is running agentic AI on systems that keep the US power grid running — and the public cannot read the rules that govern it

Idaho National Laboratory is deploying agentic AI to detect threats to critical infrastructure, preempt attacks, and quarantine systems without waiting for a human to approve the action. The lab's chief information officer, Mark Holterman, says the autonomous operations are constrained by a Community of Practice, written guardrails, and human checkpoints. When asked what a human checkpoint looks like in practice, Holterman described it as a requirement for auditable logs and traceable decisions — not a specific human approval step before an autonomous action executes. Those governance documents are not classified. They are simply not published.

The gap is real and visible. INL's cybersecurity division publishes guidance on operational technology protection, and the lab maintains a Center of Excellence for intelligent automation alongside a separate Center of Excellence for advanced analytics and LLM models, according to Federal News Network. The guardrails Holterman described are meant to govern which AI tasks are permitted to operate autonomously, what data the systems can access, and how autonomous actions are logged and reviewed. That governance framework — the existence of rules, the categories they cover — is described in broad terms by INL. The specific content of those rules is not public.

The Genesis Mission — a DOE initiative launched by executive order in November 2025 — connects 17 national laboratories through a shared AI and supercomputing platform with a stated goal of doubling the department's science output. The mission's governance documents are publicly available and describe cross-lab AI coordination at scale. They do not contain a specific published rulebook for when a national lab AI is permitted to act autonomously on operational technology — and how such decisions would be reviewed if something went wrong.

In Holterman's characterization, nation-state actors are the primary adversarial threat targeting US critical infrastructure with AI-enabled tools, and the defensive AI work is an active response to that threat environment, Federal News Network reported. The defensive response — autonomous detection and quarantine — is running in production. INL declined to say whether its guardrails contain operational specifics that would require classification to protect, or whether they are administrative in nature and simply unpublished. The rules governing that response are not public.

Pacific Northwest National Laboratory, which operates more than 500 edge compute environments across different security classifications, is part of the Genesis Mission effort. PNNL is separately building an autonomous experimentation team — informally called the "pit crew" — to support AI-driven science workflows, GovCIO Media reported. That work is largely experimental and isolated from production infrastructure. INL's agentic cybersecurity deployment is not experimental.

The practical accountability question is not abstract. If an autonomous AI at a national lab quarantines a control system and triggers a cascade failure, there is no public document that explains what the AI was permitted to do, who approved the configuration, or how the decision would be reviewed after the fact. INL says the guardrails exist and are written. Whether the lab's refusal to publish them reflects genuine security classification concerns or governance gaps — or some combination of both — is a question the public cannot answer from the outside. The lab describes guardrails. The public cannot evaluate whether those guardrails are substantive constraints or framing language.

The current moment is different from prior decades of lab oversight in one specific way: the speed and scope of autonomous action that agentic AI enables exceeds what existing public governance documentation anticipated. The Genesis Mission is explicitly designed to scale AI across the national lab system. INL's current arrangement — real autonomous systems running on critical infrastructure under unpublished rules — may be a leading indicator for how the rest of the system handles the same problem.

What to watch: whether DOE's Genesis Mission governance process eventually produces published standards for operational agentic AI at national labs — or whether INL's unpublished rulebook becomes the de facto standard as the mission scales AI coordination across all 17 laboratories.