
The second-highest alert in the NSA’s three-tier system, “Code Yellow,” was assigned to an outage on NSANet, the agency’s internal network for transmitting classified information, which it confirmed on 17 November. Apple experienced widespread disruption on the same day after an internal system update affected its employee portal, internal sign-in systems and remote VPN access. Consumer services, including iCloud, the App Store and Apple Pay, were also affected, while sales and repairs were interrupted at numerous retail stores around the world.

Delayed transaction processing at JPMorgan the previous day had stemmed from a software bug in its on-premises hardware systems. Alongside more than 3,000 outage reports recorded by Downdetector, related disruption was reported by customers of Bank of America, Wells Fargo and other institutions. A communications failure three days before that had led the FAA to order a nationwide ground stop, leaving more than 5,100 flights cancelled and nearly 10,000 delayed. These events followed the failure of core Microsoft 365 services in at least 70 countries on 12 November, which marked the start of the sequence. Outlook, Teams and Copilot all went down together, and the disruption persisted for more than two days.
Within six days, five critical systems had failed. No incident was identified as an attack or attributed to AI. Treating each as a separate technical failure without malicious intent, however, reveals a difficulty less readily resolved than an attack.
Microsoft Suffers Two-Day Outage After a “Capacity Upgrade”
Routine rerouting of user traffic was the starting point, according to the preliminary customer report Microsoft published on 15 November and Bloomberg’s reconstruction of events. Teams traffic was automatically transferred on the evening of 11 November to a different server system already processing Outlook traffic. An unrelated problem had reduced the available capacity. The combination activated a bug dormant in the company’s data centres since July, causing servers to run out of memory and crash as traffic rose sharply.
The disruption worsened when Microsoft’s automated failover system directed further traffic towards servers it regarded as healthy, allowing the fault to spread. More than 80% of Outlook servers had become unresponsive by the peak in business traffic on the morning of 12 November. Identifying the fault took Microsoft a further 15 hours; only then did deployment of a fix begin.
Laboratory testing had failed to reproduce real-world traffic patterns and hardware configurations, as the explanation from Microsoft Executive Vice President Ramanna made clear. The company’s tests usually emphasised older systems with less processing power. Yet the crash-inducing bug turned out to be “more destructive” on newer hardware, exposing a gap in the testing process.
People merit just as much consideration as technology in explaining the failure. Repeated rounds of redundancies and voluntary departures had affected the core infrastructure teams at Microsoft 365 in recent years, a user identified as a Microsoft employee claimed on the anonymous workplace forum Blind. The departure of engineers with long experience maintaining legacy systems had left remaining colleagues struggling to diagnose complicated faults at speed. Restoring systems through rollbacks could now take five hours, rather than the roughly 45 minutes previously required. At one point in this incident, internal incident-management tools reportedly failed to load as intended. Engineers were therefore forced to work late into the night across departments, investigating code and service architectures left by former colleagues.
There has been no public response from Microsoft to those claims. They do, however, fit with the disruption’s unusually long duration and the unsuccessful attempts to resolve it through rollbacks.
AI Accelerates Change as Infrastructure Renewal Falls Behind
AI was not blamed for any of the five incidents: the FAA and NSA excluded it as a factor, while Microsoft specifically rejected the involvement of “AI or agents”. That does not make “not caused by AI” equivalent to “unrelated to the AI age”.
The gap is widening between the pace of change and the renewal of the infrastructure supporting it. AI is increasing both the frequency of changes and the intensity of workloads. As traffic patterns and hardware configurations alter and software updates become more frequent, testing environments and incident-response capabilities struggle to match the pace. Infrastructure renewal follows a different timetable. Decades are the relevant timescale for the FAA’s copper cables and routers; generations shape the evolution of NSA networks. At Microsoft, meanwhile, the internal expertise available to maintain legacy systems is diminishing with each quarter.
With AI enabling system vulnerabilities to be discovered in large numbers at unprecedented speed, companies and government agencies face considerable pressure to remedy them immediately. Multiple patches are often deployed together as a result, imposing substantial demands on systems not designed for such rapid changes. The connection is not that AI generates vulnerabilities. It is that AI quickens the rate of change, whereas the systems that must accommodate it were built to operate at a slower pace.
This process is particularly clear in Microsoft’s experience: routine traffic rerouting activated a bug dormant since July, and automated failover magnified the resulting failure. Lost expertise in legacy-system maintenance then prolonged recovery to an unusual degree. AI was not an individual step in that chain, but the pressure of AI-driven cloud-services expansion formed the backdrop to every stage.