Can AI improve the very grid that it demands so much from? Anthropic’s recent Mythos announcement suggests it might.
The power grid works better than it probably should given that it was built over the last hundred years without a central blueprint. It is the product of an untold number of isolated decisions, stitched together into a system that is now facing significant load growth and sophisticated AI driven cybersecurity threats.
AI models, like the recently announced Mythos, can autonomously track down hidden software vulnerabilities at a scale and speed not seen before. Applied to a power grid running on aging software, that capability is a serious threat — but also a potentially powerful asset.
What Actually Runs the Grid
The grid is the backbone of our entire society. More than 3,000 utilities own and operate 600,000 miles of interconnected power lines that collectively underpin all sixteen of the nation’s critical infrastructure sectors as defined by the U.S. Cybersecurity and Infrastructure Security Agency.
Grid operators began relying on digital software in the 1960s to monitor and manage power flows through Supervisory Control and Data Acquisition systems, known as SCADA systems.
Since then, utilities have layered new software throughout the grid at varying paces and with little uniformity. Today many substations and energy management platforms run software written years or even decades ago.
These software tools provide visibility, telling operators where power is flowing, when lines are overloaded, and where to reroute it. Losing that visibility is like cutting the screens in an air traffic control tower. The planes are still in the air. Power is still flowing. But no one can see where everything is going or how to prevent a collision.
When the Screens Go Dark
The 2003 Northeast Blackout drove home the impact of grid software failure. A failed software alarm left grid operators with no warning of an overloading power line. Their screens reflected all clear, yet within two hours, power outages cascaded through the interconnected grid. Fifty million people across eight states and Canada lost power. The outage caused an estimated $6 billion in damage and contributed to 11 deaths. Grid operators lost visibility into their own system, and a containable incident became the largest blackout in North American history.
That was a software bug, one of many, that happened by chance. Imagine if it were exploited intentionally. That is the precipice we now stand on.
Ten Months Undetected
Adversaries have long considered the power grid a primary target for cyberattack. U.S. intelligence agencies confirm state-sponsored actors have breached critical infrastructure networks.
In April 2026, CISA issued an advisory warning that hackers have exploited SCADA systems, causing what the agency described as “operational disruption and financial loss.”
In 2024, a China-linked group known as Volt Typhoon infiltrated a Massachusetts electric utility and stayed hidden for ten months. No alarms triggered and the lights never went out. But the perpetrators had ample time to copy network diagrams and collect sensor data before seeing themselves out. According to Dragos, the security firm that investigated the breach, the Massachusetts incident is not an isolated case.
Mythos-level capabilities raise the stakes further. Better, faster, cheaper and more accessible, these tools enable anyone from nation-states to individual actors to autonomously probe legacy systems, identify precise vulnerabilities, create disruptions and engineer a 2003-type event.
Using the Threat to Beat the Threat
Ironically, the grid, like AI, scaled faster than our ability to secure it. The task of comprehensive grid security is daunting given a system that is too vast, code that is too old and ownership that is too fragmented. Now, advanced AI provides a new paradigm.
We have seen similar technology transformations. The film Hidden Figures told the true story of scores of human calculators needed in the early days of NASA spaceflight. Then computers arrived and the scale of what was possible changed overnight.
Mythos is that transition for grid security. For the first time a capability can comb through decades of layered, legacy code, across thousands of systems and find what human teams never had the time or scale to find.
That logic drives Project Glasswing, Anthropic’s initiative giving roughly 50 vetted organizations access to Mythos to hunt for vulnerabilities and patch them before they can be exploited. Energy underpins every sector on that list.
Applying that capability comprehensively across the grid is a different challenge entirely. The patchwork of utilities and governance structures makes coordinated remediation complex and expensive. It cannot be solved utility by utility, budget cycle by budget cycle.
Securing the grid is a national security imperative. Mythos just made the cost of ignoring it too great while handing us the tools to finally achieve it.
