Artificial Intelligence (AI)

Four Outages in Six Days, and No Quick Way to Tell Accident from Attack

The National Security Agency’s classified network, NSA Net, has gone offline, the fourth major American system to fail in six days. Officials call it a glitch rather than a hack, but the wider pattern of failures, compounded by AI-driven attacks that exploit unpatched systems, is raising the stakes for telling accident from intrusion.


  • Nov 18, 2026
  • 5 min read

Share this article
Four Outages in Six Days, and No Quick Way to Tell Accident from Attack

The NSA's classified network has gone down. Officials are calling it a glitch, not a hack. The pattern behind the failures is the more important story, and AI is making it more dangerous.

The National Security Agency has lost its own secret network. NSA Net, the closed system on which the agency's staff exchange classified material, went offline on Tuesday, according to six people familiar with the matter. Nothing so far points to a cyberattack. Staff have been told to arrive at staggered times while the agency works out how far the problem reaches. Backup channels mean urgent intelligence can still reach the president, who is travelling in Asia. No cause has been made public.

It is the fourth major American system to fail in six days, and the other three explain a good deal about this one.

The week everything stopped

On Thursday, Microsoft's Office 365 service went down worldwide for about a day, taking Outlook and Teams with it. The company blamed a problem during an earlier network capacity upgrade. Two days later, its threat team spotted a phishing campaign that was exploiting the disruption.

On Friday afternoon, the Federal Aviation Administration halted every departing flight in the country. Flying resumed after four hours and 45 minutes, but about 12,500 flights were cancelled around the stop. The agency says its ageing equipment became unstable when too much data flooded the network. Early findings point to two possible origins: a radio site in El Dorado, Arkansas, and a radar site in Amarillo, Texas, both of which had recently received replacement routing equipment. At El Dorado, the contractor L3Harris had been rolling out a new router. Officials describe that job as routine maintenance, and they are still working out whether an installation error, a configuration mistake or a faulty router set the problem off. The FAA sees no sign of an outside attacker.

On Monday, JPMorgan Chase customers spent much of the day unable to log in, check balances or move money. More than 4,700 outage reports had been filed by mid-afternoon. The bank called it a technical issue, not a cyber matter, and later blamed a software bug in its on-premise hardware systems. By evening, it said transactions were being processed normally. It has not said what the bug was or what set it off.

Three systems, three owners. The stories nonetheless rhyme.

What the failures share

Start with the obvious. Each of the three failures with a stated cause points inward, at the owner's own systems rather than at an intruder. In two of them, someone was changing something. Microsoft was upgrading network capacity, and the FAA's contractor was installing new equipment. Chase blames its own software but has not said what triggered it.

The FAA is partway through a vast overhaul. Congress has approved $12.5 billion, and the White House wants $30 billion more. Thousands of copper lines have already been converted to fibre. Every swap places new equipment alongside old, and that mixing is itself a source of risk. Officials call the router work routine and point to the ageing network. Yet working out which of the two was at fault is the whole purpose of the investigation. Modernisation, in other words, is risky before it becomes safe.

The damage also travels. The FAA's trouble began at one or two sites, but it flooded the network linking the agency's facilities and unsettled 17 of its 22 en-route control centres. As a precaution, it switched off a broadcast function on all its new routers, then on about 60 similar ones. Chase's outage spread in a smaller way. Wells Fargo said its online banking was temporarily unavailable, and Zelle drew a spike of complaints. Bank of America and U.S. Bank said their own systems were not directly affected.

Why the first hours are the problem

Change and spread make the first hours confusing. An accident and an intrusion can look identical from the outside: dark screens and a pledge to investigate. Only forensic work tells them apart, and that takes time. This is where the NSA stands today, which is why "not a cyberattack" is better read as a claim still being tested than as a settled finding.

Delay has always been awkward. It is now dangerous, because real attacks are close at hand. Some French supermarket fridges were pushed to temperatures of up to 50°C through Danfoss controllers that the company says were old or unpatched. A second wave in early November reached at least seven countries by exploiting a default password.

Where AI comes in

Automation shortens the clock. On Monday, Google detailed a piece of software it calls exploit_agent. It watches for newly released security patches, works out what each one repairs, and then writes and launches an attack against machines that have not yet installed the fix, with no person involved at any stage.

It feeds on a gap that defenders already struggle to close. Verizon's 2026 breach report puts the typical time to repair an exploited flaw at 43 days, up from 32. Only 26% of the entries on the federal list of known-exploited flaws have been completely fixed.

That leaves defenders in a bind. They are told to patch faster, yet two of this week's outages have been tied to upgrade or installation work. Patch quickly and you risk an outage; patch slowly and you risk an exploit.

AI can help the defenders as well. Anthropic says its Project Glasswing verified more than 129,000 flaws between April and July. But finding holes is now cheap for everyone, while repairing them still requires people, budgets and downtime. Discovering a flaw is also not the same as seeing it exploited. VulnCheck found that only 2 of 300 vulnerabilities discovered by Anthropic or Glasswing had been exploited in the wild. The bottleneck is fixing, not finding, and the queue is growing faster than crews can clear it.

Who is betting on the fix

Wall Street has noticed. Anthropic is due to fix its IPO price, with first trades expected on Thursday. Palo Alto Networks reports after the market closes tomorrow, and CrowdStrike reports on 1 December. Investors are wagering that this trap will turn into budgets.

What happens next

On the evidence so far, this looks more like accident than attack, though that is a judgment rather than a finding. Microsoft, the FAA and Chase have each pointed to internal causes and ruled out an attack, and none has produced an intruder.

Four explanations remain outstanding: Microsoft's full report, the FAA's root cause, the nature of Chase's bug and what triggered it, and whatever took NSA Net offline. If any one of them breaks the pattern, an attack will have spent its first hours looking like bad luck.

The test is no longer only what broke. It is how fast anyone can tell.


Download Research Report
Please enter a valid phone number