Anthropic brings Claude to power plants and water facilities, but cyberdefense's bottleneck remains physical

By BankInfoSecurity · October 9, 2026. Anthropic has launched the Critical Infrastructure Defense Program, an initiative that deploys its latest-generation Claude models — along with engineers on the ground and threat research teams — at 11 companies that advise operators of power grids, water plants, factories, transport networks and government systems. The founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation, and they are already using Claude to fix vulnerabilities and pass those fixes on to their clients, according to Anthropic.

The program comes after Anthropic expanded its Cyber Verification Program to all critical infrastructure operators, giving them access to models with fewer safety barriers — what the company calls "reduced guardrails" — for both defensive and offensive cybersecurity tasks. In practice it is a double bet: selling controlled offensive capability to those defending the power grid, and now adding specialized human hands so that capability translates into real patches.

As sector context, the urgency is not rhetorical. In August, a small power plant in the United Kingdom was out of service for four days. In July, more than 30 community water systems in Minnesota had to switch to manual control for two days after a coordinated attack. In both incidents, programmable logic controllers — PLCs, the devices that automate industrial machinery — have been pointed to, amid warnings about Iran-linked actors targeting internet-exposed controllers, the report notes.

The underlying technical problem, explains John Gallagher, vice president of OT cyber hygiene firm Viakoo, is that many OT systems (operational technology, which controls physical processes such as turbines or water pumps) are only patched during quarterly maintenance windows, while AI-driven threats have cut the time to exploitation to one or two days. Gallagher warns of a specific risk: "vulnerability fatigue." Announcing a critical flaw faster protects nothing if applying the patch remains as hard as before.

Jen Sovada, general manager for the public sector at OT security firm Claroty, sums it up this way: the challenge is not finding risks, but deciding which ones demand immediate attention when shutting down a system can have serious consequences. "AI cannot replace the human expertise needed to understand how these systems operate and what is at stake if something goes wrong," she told ISMG, BankInfoSecurity's editorial network. Josh Corman, of the Institute for Security and Technology, adds another layer: conventional IT cybersecurity best practices can be dangerous in "fragile" OT environments, where a simple port scan can reboot or break an old controller. Much of that equipment — from Schneider, Rockwell, Honeywell, Eaton or Siemens — has a service life of 30 years or more and, according to Corman, sometimes allows only one patching window a year, and with luck somewhat more often.

Our take is that this program accurately captures the asymmetry that defines critical infrastructure cybersecurity in 2026: attacking has become faster thanks to AI, but defending is still limited by the physics of equipment that cannot be switched off or updated at will. Claude's real value here is not in finding vulnerabilities — existing commercial tools already do that reasonably well — but in triage: saying which of a thousand alerts truly matters before someone mistakenly disconnects a pump at a water treatment plant. It is the same lesson we draw from the cybersecurity benchmarks we follow regularly: what matters is not raw capability on already-saturated tests, but performance on expert tasks without assistance, the ones that really separate a useful tool from one that only generates noise.

There is also a market-power reading. Anthropic is positioning itself as a technical partner to the world's largest security consultancies — Accenture, Deloitte, PwC, Booz Allen — and to the most established OT specialists — Dragos, Nozomi Networks, Rockwell Automation — reinforcing its role as strategic infrastructure and not merely a model provider. But, as Anthropic itself acknowledges, the program starts with "a small group of trusted vendors" covering only a fraction of the country's critical infrastructure. Smaller operators — rural electric cooperatives or municipal water systems — are predictably not in that first wave. If AI-assisted security reaches those who can already afford a top-tier consultancy first, the resilience gap between large and small operators could widen before it closes.

In the short term, the honest risk is that the program amounts to a gesture if remediation capacity does not grow at the pace of detection: that is precisely the "vulnerability fatigue" Gallagher warns about. But if AI manages to establish itself as the filter that separates the urgent alarm from the noise in systems that cannot be switched off to think it over calmly, we will be looking at a small but concrete brick in that future of invisible, reliable infrastructure that underpins our core thesis: a society where power, water and transport simply work, freeing human operators to deal with what truly requires their judgment rather than chasing alerts around the clock.