Analysis · Data & AI Infrastructure
Every Outage This Week Began With a Change, and AI Is Multiplying the Changes
Microsoft, the FAA, Chase, Apple and the NSA all pointed to updates, installations or maintenance. As AI multiplies patches, code and agent actions, change management becomes one of the most important controls enterprises have.
By The AI Enterprise Magazine Team6 min readAnalysis · Weekly briefing

Key takeaways
- Change, not attack. Stated causes include a July update bug, a router install, an internal update and maintenance.
- Three accelerants. AI-found vulnerabilities, AI-written code and agents with operational authority all raise change volume.
- Controls that failed. Test hardware that did not match production, latent defects and uncontained blast radius.
- Safer, not slower. Progressive delivery, automated rollback and agent-attributed changes keep pace without freezing.
Read the official explanations for the outages that hit the United States in mid-November and a single word keeps appearing in different forms. Microsoft described a bug from a July update that surfaced after a network change. The FAA described a router installed during a copper-to-fiber conversion. JPMorgan Chase described a software bug. Apple said "an internal update caused some outages." The NSA said its network degraded during scheduled maintenance. The Director of National Intelligence summed it up as "an upgrade in patching and upgrade in systems."
None of these organizations was attacked, according to everything they and the government have said. Each of them changed something, and the change did not behave as expected. That is not a new story in information technology. Anyone who has run production systems knows how often serious incidents trace back to a change rather than to a random hardware failure. What is new is the pace at which change is arriving, and the degree to which AI is driving it.
Whether or not AI touched any of the systems that failed last week, the lesson for enterprise leaders is the same. Change management, long one of the least glamorous disciplines in IT, is becoming one of the most important controls an organization has over its AI future.
How did each outage start with a change?
The details matter, because they show how ordinary changes become extraordinary failures.
At Microsoft, according to Implicator's account of the company's preliminary report, a defect had been sitting in production since a July update. It did nothing until November 12, when extra traffic arrived while some capacity was offline, and more than 80% of the company's networking gateway servers became unresponsive. Microsoft's testing had missed the problem because its lab hardware was weaker than production hardware. The change that caused the outage was months old, and the change that exposed it was routine.
At the FAA, a contractor installed a new router at a radio site in El Dorado, Arkansas, as part of replacing old copper lines with fiber. Hewlett Packard Enterprise, which owns the router maker Juniper, said the device "functioned as it was configured to by the installer," Implicator reported. The FAA said its network was flooded with roughly 1,000 connection attempts every four to five seconds, according to TechRepublic, and 17 of its 22 en route centers lost stable communications. Transportation Secretary Sean Duffy said a separate router in Amarillo, Texas, had been upgraded weeks earlier and also misbehaved, Fortune reported. A modernization project, the kind the FAA has been urged for years to accelerate, was the trigger.
At Apple, an internal update. At Chase, a bug in an on-premise hardware system. At the NSA, maintenance that Implicator, citing remarks by Director of National Intelligence Jay Clayton, linked to a patch for vulnerabilities disclosed the month before. Some of these accounts are thin, and investigations are still under way. But not one organization has described a failure that arrived without a change preceding it.
Why does the pace of change matter now?
For most of the history of enterprise IT, the rate of change was limited by how fast people could write, review, test and deploy it. Three developments in 2026 have loosened those limits at once.
The first is vulnerability discovery. Microsoft fixed 1,093 vulnerabilities in its November security release and 1,407 in October, according to Tenable, compared with a record of 167 just a year earlier. Microsoft has said its own AI scanning systems are part of the reason. Every one of those fixes is a change that someone has to deploy.
The second is AI-written code. Coding agents now generate and modify a growing share of enterprise software, and they do it faster than human reviewers were staffed to inspect. The same week as the outages, Microsoft put new coding models into GitHub Copilot and OpenAI released new models for its Codex coding agent. Each of those tools increases the amount of change an engineering organization can produce in a day.
The third is agents with operational authority. Vendors are now selling agents that run continuously inside a company's systems and take actions on their own. Microsoft's Autopilot, introduced at its Ignite conference on November 17, is described as having "its own identity, memory, and tools," according to Microsoft. Agents that act on infrastructure are, by definition, agents that make changes.
Why didn't existing controls catch these failures?
The outages point to three weaknesses that become more dangerous as change accelerates.
Testing environments did not match production. Microsoft's lab missed its bug because the hardware was different. That is a common compromise, since full production replicas are expensive, but it means the most important class of defects, the ones that appear only at real scale and load, are the least likely to be caught before deployment.
Latent defects were invisible. A bug introduced in July and triggered in November means that a change can pass every test, run quietly for months and then fail under conditions no one anticipated. When the number of changes in production multiplies, so does the number of latent defects waiting for the right trigger.
Blast radius was not contained. The FAA's problem began at one or two sites and spread across most of the national network. Microsoft's spread across regions. In tightly connected systems, a change in one place can take down services far away, and the controls that should limit that spread, such as segmentation, rate limiting and automatic isolation, are often added after the fact.
None of these weaknesses is caused by AI. All of them become more costly when AI multiplies the number of changes flowing through a system.
What does good change management look like at AI speed?
The traditional answer to risky change was to slow it down with approval boards, freeze windows and manual sign-offs. That answer does not survive an environment where the volume of necessary security fixes has risen nearly tenfold in a year. Organizations that slow down will leave known vulnerabilities open to attackers who have the same AI tools. The goal has to be making change safer without making it slower.
Several practices point the way. Progressive delivery, in which a change reaches a small share of systems or users first and expands only if health signals stay normal, limits the damage any single change can do. Automated rollback, triggered by the same health signals, shortens the time a bad change stays live. Production-like testing for high-risk components, even if it is expensive, addresses the class of defect that hit Microsoft. And clear tiering of systems by criticality lets organizations apply heavier controls where an outage would be most costly.
There is also a role for AI on the defensive side. If agents are generating changes at machine speed, organizations will need automated systems to assess the risk of each change, predict its blast radius and watch for anomalies after deployment. The irony is that the best tools for managing AI-accelerated change may themselves be AI systems, which then need their own governance.
What should enterprise data and AI leaders do now?
Data and AI leaders are not usually the owners of change management. That is part of the problem. Decisions about adopting coding agents, connecting AI to vulnerability scanning or giving agents operational permissions are often made in AI programs, while the processes that absorb the resulting changes sit in infrastructure and operations teams.
The first step is to connect the two. Every AI initiative that will increase the rate of change in production systems should have an explicit conversation with the teams that own change and incident management about capacity, testing and rollback. The second is to measure. Change failure rate, time to restore service and the volume of changes per week are long-standing measures of software delivery performance, and they should be tracked alongside AI adoption metrics, so leaders can see whether faster change is also making systems less stable. The third is to treat agents as a new category of change author. An agent that modifies code or infrastructure should be subject to the same review, staging and audit requirements as a human engineer, and its changes should be identifiable as its own.
Last week's outages have ordinary explanations, and investigators may well confirm them. That should not be reassuring. Ordinary failures of change are exactly the kind of failure that will multiply as AI accelerates the rate of change, and the organizations that adapt their change processes now will be the ones that are not explaining themselves next time.
Frequently asked questions
What caused the November 2026 outages?
The organizations involved attributed them to changes: a July update bug at Microsoft, a router installation at the FAA, an internal update at Apple, a software bug at Chase and maintenance at the NSA.
How does AI increase outage risk?
AI raises the volume of change through faster vulnerability discovery, AI-written code and agents that act on systems. More changes mean more chances for defects to reach production.
How should enterprises manage AI-driven change?
Use progressive rollouts, automated rollback and production-like testing for critical systems, track change failure rates, and require agents to make attributable, reviewable changes.
Sources
- Implicator.ai: Outages Rattle Americans as Spy Chief Blames Patching
- TechRepublic: FAA Says Its New AI Didn't Ground US Flights — So What Does SMART Actually Control?
- Fortune: Did a cyberattack hit the FAA? Probably not, transportation secretary says
- Tenable: Microsoft's November 2026 Patch Tuesday addresses 1093 CVEs
- Microsoft Copilot Blog: Copilot at Microsoft Ignite: Expanding what’s possible with intelligence built for work
- Fox News: Tech outages live updates: Microsoft, banks and FAA
Filed under: Data & AI Infrastructure, AI Governance & Trust, Agentic AI
Become a Sponsor







