Four Outages in Six Days: FAA, Microsoft 365, Chase and the NSA
Microsoft 365, the FAA, Chase and the NSA's classified network all failed between November 12 and 17, 2026. What happened in each, the technical causes as currently stated, what's still disputed, whether it was a cyberattack, and the change management lessons for defenders.
Watch the video version on YouTubeIn this deep dive
- Microsoft 365: a latent July bug, a Teams failover and 14% lost capacity, then rerouting that spread the failure to 80%+ of gateway servers
- The FAA ground stop: two possible sources on the L3Harris-run FTI network, 1,000 connection attempts every 4 to 5 seconds, and what's still disputed
- Chase: a bug in an on-premise hardware system, and why Downdetector lit up for banks that weren't down
- NSANet: degraded during scheduled maintenance, and why the NSA staggered logins
- Was it a coordinated cyberattack? What the evidence says and doesn't
- Three lessons, what it means for Security+, CySA+, CISSP and CCSP, and three things to do this week
Sources: Microsoft's preliminary post-incident review (Nov 15, 2026); FAA statements and Q&A; HPE, Senate Commerce Committee and Transportation Department statements; Chase, Bank of America, U.S. Bank and Zelle statements; NSA and White House statements as reported by CNN, CBS News, Nextgov/FCW and others. Facts current as of November 17, 2026. The full write-up, with every source, is on the Skillthropic blog.
Links
Transcript
Between Thursday, November 12th and Tuesday, November 17th, four critical systems failed. Microsoft 365 went down for most of a day. The FAA ordered a nationwide ground stop. Chase customers couldn't log in or move money for hours. And the NSA's classified internal network was degraded. Four failures in six days, and plenty of people online have connected the dots and called it a coordinated cyberattack. The evidence so far says otherwise. Each one has its own cause, and at least three of the four trace back to a change someone made to a working system. We're going to walk through each outage: what happened, the technical cause as it's currently stated, what's still disputed, and what it teaches anyone who runs, defends or studies production systems. Keep in mind this is a developing story. Facts are current as of November 17th.
Let's start with Microsoft. At 16:20 UTC on Thursday, November 12th, that's 11:20 in the morning Eastern, Microsoft opened an incident for Exchange Online. Many users couldn't reach their mailboxes, and mail delivery stalled. Exchange took the main hit, and the list of affected services grew to include SharePoint, OneDrive, Purview, Intune and more. It started in the Americas and spread to every region. Microsoft declared it over just under twenty-four hours later. What makes this one worth studying is that Microsoft published a preliminary post-incident review on November 15th, and it's unusually candid. It describes three factors that together took down the networking gateway that fronts many Microsoft 365 services.
First off, a latent defect from July. A change to the gateway servers introduced lock contention inside an internal Windows process under certain load. In very simple terms, lock contention is when many threads compete for the same lock, so most of them spend their time waiting instead of working. It passed testing because the problem is worse on newer machines with more cores, and the load tests ran on weaker hardware with fewer cores and a different traffic mix. Then it ran worldwide for months without trouble. Second, extra load. At 01:07 UTC, about fifteen hours before the outage, an automated failover moved Teams traffic onto gateway servers shared with Exchange Online. And then lastly, reduced capacity. Nearly fourteen percent of the gateway's servers were out of service from an unrelated issue the day before. On its own, that was within the service's headroom. All three together, during peak US traffic, pushed the servers into exactly the conditions that triggered the July bug.
Here's the part I want you to remember. Requests piled up until servers stopped responding. Then the system's own resilience made it worse. Automatic rerouting sent each failed server's traffic to the healthy ones, which pushed them over too. In under two hours the failure had spread across the fleet, and at the worst point more than eighty percent of the gateway servers were unresponsive. The bulkheads that normally keep a failure in one region from spreading didn't hold, because of a separate issue. Servers Microsoft restored that evening stayed healthy at first, then fell over again once traffic reached them. And because the defect had shipped months earlier, the normal rollback tools, the staged release rings and the feature flags couldn't quickly revert it. The eventual fix was a code revert, first applied early on November 13th. One more detail worth knowing: Microsoft's explanation changed. On November 13th it pointed to an issue during an earlier network capacity upgrade. The review describes a code revert of a July change instead. The final version is due November 20th. Microsoft says it found no indication of malicious or cyber activity.
What does it teach? This is the textbook shape of a large outage. No single cause. A latent defect, a load shift and a capacity shortfall lining up at once, and then a failover mechanism spreading the damage. Two lessons stand out. A test lab that doesn't match production hides bugs. And rollback tooling is built for recent changes. A defect that sits quietly for four months has aged out of it, so you need a way to revert old changes too.
Next, the FAA. Shortly after 3:30 in the afternoon Eastern on Friday the 13th, radar and radio data feeding US air traffic control became unstable. At 3:54 the FAA ordered a nationwide ground stop, and it lifted it at 8:15, about four hours and twenty minutes later. The instability hit seventeen of the twenty-two Air Route Traffic Control Centers, the facilities that handle aircraft at cruising altitude, plus several approach control facilities. The FAA stresses this was not a full outage. Some frequencies and radar feeds stayed reliable. About three thousand five hundred aircraft were in the air when it started, and all of them landed safely. The disruption rolled through the weekend, with thousands of cancellations on Friday, Saturday and Sunday.
The FAA's voice and data circuits run over a network called FAA Telecommunications Infrastructure, or FTI, owned and operated by the contractor L3Harris. L3Harris is converting that network from copper to fiber and installing new Juniper routers as it goes. The FAA has named two possible sources of the trouble: a radar site in Amarillo, Texas, whose router was replaced about three weeks earlier, and a radio site in El Dorado, Arkansas, with a newly installed router. According to the FAA, those routers flooded the network at a rate of a thousand connection attempts every four to five seconds, enough to destabilize the circuits carrying radar and radio data. The network stabilized after L3Harris blocked the two sites. The FAA has since disabled a broadcast function on more than sixty other Juniper routers and paused new installs, and sites stay on their older Cisco routers in the meantime. For scale, Senate Commerce Committee Democrats say three thousand six hundred forty-eight routers are due to be replaced, about four hundred have been installed, and only sixty-nine are operational.
What's disputed? Almost everything about why. Transportation Secretary Sean Duffy initially blamed aging air traffic control technology and overloaded routers, and later asked whether it was an installation error by L3Harris or a faulty router. Senate Democrats point out that the routers implicated are the new ones. The FAA's own Q&A points at instability in the legacy system. HPE, which owns Juniper, says its device did not fail or act improperly, and functioned as it was configured to by the installer. In a letter, thirteen Senate Commerce Democrats asked for a pause on installs, a technical briefing and written answers by November 23rd, noting that identifying the device that failed doesn't establish why it failed. On the cyber question, the FAA says there's no indication of a cyberattack or an outside threat, and FAA Administrator Bryan Bedford said there's zero evidence of anything nefarious. The routers are being tested at an L3Harris lab in Florida with FAA staff and FBI agents present.
What does it teach? A device flooding a network with connection attempts or broadcasts is a classic failure mode, and networks have controls for it. Storm control caps broadcast and multicast traffic on a port and drops it above the limit. Control-plane rate limiting protects the router's own brain. And segmentation keeps one site's noise from reaching every center. The bigger lesson is about migrations. During a copper-to-fiber transition, the network is a hybrid of old and new, and new equipment talking to old equipment is exactly where untested behavior lives. HPE's point is the reminder: configuration is part of the change. This isn't the FAA's first change-driven ground stop, either. In January of 2023, the nationwide NOTAM outage was traced to contract staff who unintentionally deleted files while fixing a database sync. Not an attack. A change.
On Monday, November 16th, it was Chase. Customers reported problems with logins, balances, transfers, withdrawals and Zelle payments for hours. Downdetector reports for Chase climbed through the day, to around five thousand in the early afternoon. Chase said customers still had access to their accounts and funds, though some transactions were slow, and by about 6:30 in the evening Eastern, transactions were posting normally. Chase attributed it to a temporary issue related to a software bug in an on-premise hardware system. It said flatly that the problem was not cyber related, and that it was not a systemic issue. Chase hasn't named the system or the vendor. An unnamed FBI official told Fox Business the disruption traced back to a software release the bank rolled out earlier that Monday, and that the impact was limited to Chase. Chase itself hasn't confirmed that.
Here's the interesting part. Downdetector reports also spiked for Bank of America, U.S. Bank and Wells Fargo. Were they down? It doesn't look that way. Chase said other banks were affected only to the extent they were a party to a transaction with Chase. U.S. Bank said it wasn't directly impacted. Bank of America said its systems operated normally throughout the day. And Zelle said Zelle is not down. Zelle carries the payment instructions, but the money moves between bank accounts, so when the biggest participant stalls, payments stall for everyone dealing with it. If your payment to a Chase customer hangs, you report your own bank as down. The lesson: crowd-sourced outage data measures symptoms, not causes. Incident responders have to separate where users feel the pain from where the fault is, and the same discipline applies when you're reading Downdetector during your own incident.
Then on Tuesday, November 17th, CNN, CBS News, Nextgov and others reported that NSANet, the NSA's classified internal network and its primary platform for top-secret communications, was disrupted. Some staff couldn't reach their workplace systems. According to CBS sources, the agency staggered arrival times for non-emergency staff at Fort Meade so everyone wouldn't log on at once. An NSA spokesperson said the agency experienced temporary degradation during scheduled routine maintenance, that some internal systems were affected, and that mission-critical capabilities remain fully operational. Everything beyond that comes from anonymous sources. Two told CBS it was routine system patching. Four of six told Nextgov it may be linked to a series of software upgrades. As of Tuesday evening, restoration was still ongoing. On the cyber question, the wording is no evidence, not ruled out. A White House spokesperson said the administration sees no evidence it was a cyberattack.
What does it teach? The NSA is subject to the same rule as everyone else. Maintenance is a change, and changes break things. Classified networks are usually air-gapped from the internet, which limits outside threats, but it also limits your options during recovery. You can't fall back to a cloud service or pull a fix from a vendor in minutes. And the staggered arrivals are a practical recovery technique. When a system comes back, a thundering herd of simultaneous logins can knock it right back over, which is exactly what returning traffic did to Microsoft's restored servers.
So, were the four connected? Nothing published so far supports that. Homeland Security Secretary Markwayne Mullin called them individually unrelated. Each organization has given a non-malicious cause, and the causes are different in kind: a code defect at Microsoft, router behavior on an FAA contractor's network, a software problem in Chase's own systems, and maintenance at the NSA. They also span unrelated infrastructure: a global SaaS platform, a contractor-run telecom network, a bank's own data centers and an air-gapped classified network. A single attacker would need a foothold in all four. Two caveats. First, no evidence of a cyberattack isn't the same as ruled out, and the FAA's root cause is still under investigation, with the FBI involved. Second, early explanations have already shifted. Microsoft's did. So the final reports matter. But large systems fail all the time. A cluster of failures in one week is usually coincidence plus heightened attention, not a campaign.
Strip away the headlines and three lessons are left. First off, every change is a risk, including the routine ones. Patching, upgrades and hardware swaps caused or set up most of these outages. Change management isn't paperwork. It's impact analysis, a tested backout plan, and a maintenance window sized to the risk. This is not a checkbox activity. Second, contain the blast radius. Bulkheads, segmentation, storm control and rate limiting exist so one bad component becomes a local incident. When they fail, so does everything behind them, and failover can carry the fault to the next healthy server. And then lastly, plan for the restart, not just the failure. Returning traffic knocked Microsoft's restored servers over again, and the NSA staggered logins to avoid the same thing.
If you're studying for a certification, this week was a gift. For Security+, availability is one third of the CIA triad, and this was an availability case study. Change management, with approvals, impact analysis, backout plans, maintenance windows and testing, and resilience, with high availability, capacity planning and recovery, are both in the objectives. For CySA+, the skill on display is triage: separating the fault domain from the symptoms, and deciding quickly whether an event is malicious. Microsoft's review is a good model of a lessons-learned report. And for CISSP and CCSP, it's business continuity, disaster recovery and cloud resilience. Microsoft's outage is a shared responsibility question for every customer. What's your plan when your SaaS provider is down for a day? And the FAA's is a third-party and supply chain risk question.
Three things you can do this week. First off, audit your last ten changes for a real backout plan. For each one, ask whether you could revert it in under an hour, including changes that went live months ago. If the answer is no, fix the plan before the next change. Second, write down your plan for a twenty-four hour email outage. Decide now how your team communicates, who declares the incident, and where the out-of-band contact list lives. And then lastly, check that one noisy device can't take down a site. Confirm storm control and control-plane rate limiting are enabled on your switches and routers, and that a new device is staged and monitored before it joins production.
Four failures, four causes, and one common thread. None of them needed an attacker. They needed a change, a bit of bad luck, and a safety net that didn't hold. Test the plan. Don't plan the test. The full write-up, with the timeline, what's still disputed and every source, is on skillthropic.com. While you're there, have a look at the study guides for Security+, CySA+, CISSP and CCSP.
Enjoyed this? Get the AI security news that matters.
Join The AI Security Brief for the top AI security news, plus what's important to the C-suite. Free, straight to your inbox.
No spam. Unsubscribe anytime.
Follow Skillthropic
Keep learning between study sessions
The same exam-focused teaching, in smaller pieces, wherever you scroll.