
On November 17, the NSA confirmed an outage affecting NSANet, its internal network for transmitting classified information. The incident was designated “Code Yellow,” the second-highest level in its three-tier alert system. That same day, an internal system update at Apple caused widespread disruptions to its employee portal, internal sign-in systems, remote VPN access, and consumer services including iCloud, the App Store, and Apple Pay. Sales and repair operations at numerous retail stores worldwide were interrupted.
A day earlier, a software bug in JPMorgan’s on-premises hardware systems delayed transaction processing. Downdetector recorded more than 3,000 outage reports, while customers of Bank of America, Wells Fargo, and other institutions also reported related disruptions. Three days before that, the FAA issued a nationwide ground stop following a communications failure, with more than 5,100 flights cancelled and nearly 10,000 delayed. The sequence began on November 12, when core Microsoft 365 services went down in at least 70 countries. Outlook, Teams, and Copilot failed simultaneously, with disruptions stretching over more than two days.

Six days. Five critical systems. None of the incidents was identified as an attack, and none was attributed to AI. Yet this very narrative—that each was an isolated, non-malicious technical failure—exposes a problem that is harder to fix than an attack.
Microsoft: How a “Capacity Upgrade” Turned Into a Two-Day Outage
According to Microsoft’s preliminary customer report released on November 15 and Bloomberg’s reconstruction of events, the outage began with a routine redirection of user traffic. On the evening of November 11, Teams traffic was automatically routed to another server system handling Outlook traffic. Combined with capacity already lost to an unrelated issue, this triggered a bug that had remained dormant in the company’s data centres since July: servers exhausted their memory as traffic surged and began crashing.
More damagingly, Microsoft’s automated failover system redirected additional traffic to other servers it considered healthy, spreading the failure. By the morning business traffic peak on November 12, more than 80% of Outlook servers were unresponsive. Microsoft then spent another 15 hours identifying the problem before beginning to deploy a fix.
Microsoft Executive Vice President Ramanna’s explanation pointed to a blind spot in the company’s testing process: traffic patterns and hardware configurations simulated in its laboratories did not reflect real-world conditions. Testing typically focused on older, less powerful systems, while the bug that triggered the crashes proved “more destructive” on newer hardware.
But the human factor deserves as much attention as the technical explanation. On the anonymous workplace forum Blind, a user identified as a Microsoft employee claimed that Microsoft 365’s core infrastructure teams had undergone multiple rounds of layoffs and voluntary departures in recent years. Engineers who had long maintained legacy systems had left, leaving the remaining staff struggling to diagnose complex failures quickly. System rollbacks that once took around 45 minutes could now take five hours. During this incident, internal incident-management tools reportedly failed to load properly at one point, forcing engineers to work across departments late into the night to investigate code and service architectures inherited from former colleagues.
Microsoft has not publicly responded to these claims. However, they are consistent with the unusually long disruption and the failure of rollback attempts to resolve the issue.
What Is Really Happening: A Widening Gap Between AI’s Pace and Infrastructure Renewal
None of these five incidents was attributed to AI. Microsoft explicitly denied that “AI or agents” were involved, while the FAA and NSA also ruled out AI as a factor. But “not caused by AI” and “unrelated to the AI age” are two different things.
What is unfolding is a divergence between two curves. One represents the frequency of changes and the intensity of workloads driven by AI: traffic patterns are changing, hardware configurations are changing, and software updates are accelerating, while testing environments and incident-response capabilities are struggling to keep up. The other represents infrastructure renewal cycles: the FAA’s copper cables and routers operate on timelines measured in decades, the NSA’s networks evolve across generations, and Microsoft’s internal capacity to maintain legacy systems is eroding quarter by quarter.
AI makes it possible to identify large numbers of system vulnerabilities at unprecedented speed. This places enormous pressure on companies and government agencies to fix them immediately, often by deploying multiple patches at once, putting considerable strain on systems never designed to accommodate such rapid change. The causal link here is not that AI creates vulnerabilities, but that AI accelerates the pace of change while the systems absorbing those changes were designed for a slower rhythm.
Microsoft’s case illustrates this mechanism most clearly. The bug had lain dormant since July. A routine traffic redirection triggered it. Automated failover amplified it. And the loss of expertise in maintaining legacy systems made recovery unusually slow. None of these individual steps was AI, yet every one occurred under the pressure of AI-driven expansion in cloud services.