The Call Is Coming From Inside the Server Room

Chase, the FAA, and the NSA all went dark within a week, and no hackers were involved. Routine maintenance did the damage. Can AI catch our own mistakes before they spread?

Three of America’s most critical systems failed within five days. None of them was hacked.

On Friday, the FAA ordered a nationwide ground stop for several hours, forcing airlines to delay and cancel thousands of flights. A router caused it. During a copper-to-fiber swap, a replacement unit began flooding other routers in the air traffic control system with more than 1,000 connection attempts every four to five seconds. The FAA said instability hit 17 of its 22 Air Route Traffic Control Centers, and it spread into several Terminal Radar Approach Controls. 

On Monday, JPMorgan Chase blamed a software bug in “on-premise hardware systems.” Chase is the nation’s largest bank, with more than 86 million customers. The bug slowed online banking, Zelle, and even transactions inside branches. The trouble spread beyond Chase. Bank of America and Wells Fargo customers got stuck when they sent money to or from Chase, and Plaid reported errors connecting Chase accounts to other apps.

On Tuesday, the NSA’s classified internal network, NSANet, went down. The agency said it “experienced temporary degradation during scheduled routine maintenance.” Sources told CNN there was currently no evidence the outage was the result of a cyberattack. 

Add last week’s Microsoft 365 outage, which reached users in more than 70 countries, and a pattern appears. The biggest threat to critical infrastructure may not be a foreign adversary or a ransomware gang. It may be the next upgrade.

We Were Warned

The FAA failure stings most because people saw it coming. Federal auditors have flagged the network as an outage risk for more than 15 years. After a 2009 router failure on the same system, an independent panel the FAA convened recommended automated tools for router configuration changes and a model of the network to test changes and upgrades. 

A model of the network that tests changes before they go live is a digital twin. That is exactly the kind of work modern AI is getting good at.

Can AI Get There First?

There’s a strong case for optimism. AI can take in years of configuration history, change logs, and incident reports, then flag a risky update before a technician touches a cable. It can simulate how one faulty router would spread trouble across 22 control centers. It can spot a flood of connection attempts in seconds rather than hours and cut off the bad device before the whole system goes down.

There’s an equally strong case for skepticism. AI is also speeding up the pace of change. More code shipped faster, much of it machine-written, means more updates that can go wrong. The Federal Reserve already worries that AI agents could set off bank withdrawals that unfold in minutes. An AI watchdog that misreads a situation could make a failure worse as easily as it could stop one.

And the FAA never lacked ideas. Its fix was written down in 2010. What it lacked was the money, the urgency, and the will to act.

That is the real question. AI can model our systems, read our logs, and warn us. Will anyone listen this time?

Related Articles

Sources

Next
Next

Better Dinosaurs: AI, Sentience, and the Catch-22 We Built