AI benefits/risks

The Microsoft and FAA incidents show how AI accelerates change in systems we can’t afford to break

A futuristic air traffic control tower with an AI managing autonomous aircraft.

COMMENTARY: AI has changed how quickly we can now write, revise, and prepare software for release. Coding agents can draft code, update dependencies and propose changes across repositories in minutes. Yet the systems receiving those changes are built from layers of services, legacy infrastructure, and dependencies that are difficult to see in full.

That gap between the speed of change and the ability to understand its effects has become a central operational risk.

[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]

Two infrastructure outages last week show how a technical change can ripple through complex systems. On Thursday, November 12, a global Microsoft 365 outage took Outlook, Teams, Copilot, and other services offline for more than 24 hours, affecting organizations in more than 70 countries. Microsoft attributed the incident to an issue introduced during an earlier network capacity upgrade and explicitly ruled out a cyberattack or AI involvement.

The following day, the FAA halted departures nationwide for roughly four to five hours after communications problems affected 17 of 22 of its regional Air Route Traffic Control Centers after instability in its telecommunications network disrupted some aircraft communications and radar feeds. The FAA said its initial investigation traced the problem to data flooding the network during work at a remote site, where a contractor was replacing copper lines with fiber and deploying a router. The investigation continues.

Neither outage was a cyberattack. Both were infrastructure-change failures — a network capacity upgrade at Microsoft, a copper-to-fiber upgrade at the FAA — that cascaded through complex, interdependent systems. The common thread: the accelerating pace of change flowing through systems that mix new and legacy components, and the verification gap that opens when change outpaces our ability to understand its impact.

New technology rarely replaces the old all at once. It arrives in layers: new services connected to existing applications, modern interfaces built over aging infrastructure, automated workflows spanning systems designed decades apart. Most businesses must modernize while continuing to operate. The difficult work is understanding what happens where those layers meet.

The FAA's November 13 ground stop brought that challenge into view. A change in one part of a complex system can have consequences far beyond it. Enterprises face that problem whenever they modernize systems that support mission-critical applications. Few have the option of starting over.

Last week’s Microsoft outage illustrates the same pattern at cloud scale. An issue introduced during an earlier network capacity upgrade surfaced after a subsequent network change, causing front-end servers to fail automated health checks. As servers were removed and clients reconnected, each reconnection triggered the exact operation that was already overloaded, creating a self-sustaining cascade — what Microsoft described as a metastable failure. The trigger was removed, but the failure persisted, because the system’s own recovery mechanisms kept amplifying the load. It’s the same issue: in complex, interdependent systems, a single change can cascade far beyond its intended scope.

AI accelerates the work entering those environments. It does not automatically accelerate our understanding of them. Coding agents can help teams build features, update dependencies and prepare fixes. GitHub reports that more than 40% of pull requests began with an agent’s first draft by September 2026, up from 17% in March. This does not mean every change reaches production faster, but it shows how quickly development practices shift. The resulting code still enters environments shaped by years of dependencies, architectural decisions, and operating constraints.

Attackers now leverage AI to build malware and automated campaigns, which increases the pressure on companies to deploy security patches from weeks to mere days. Organizations must patch more often and more quickly, increasing the volume of changes flowing through their systems — and as both the Microsoft and FAA incidents showed, it’s the change itself, not just its security intent, that can break a complex system.

As code becomes easier to produce, it’s become much harder to decide what’s ready to release. Teams must determine whether a change preserves expected behavior, introduces a security weakness, or affects a service elsewhere in the business. They also need to know how to recover if it fails. Functional tests answer some of these questions. Others require security analysis, knowledge of the environment, and engineering judgment.

The capacity to verify changes must grow alongside the capacity to generate them. Otherwise, organizations risk moving the bottleneck from development into review or allowing changes through with insufficient scrutiny. New technology must work within the architecture, dependencies, and operating limits already in place. As AI accelerates development, organizations need to understand how the changes it produces will behave within those constraints.

Context matters.

The same change can carry different risks depending on the system it enters, the data it touches, the privileges it inherits, and the services that depend on it. Verification must account for those relationships, including the constraints of legacy components and the potential impact on mission-critical applications. Assessing code in isolation cannot deliver that full picture.

Agentic development adds another layer to this environment. Agents become participants in established workflows, using development tools, accessing repositories, and proposing changes to existing applications. Securing their work requires understanding both what they produce and what they are authorized to do. Organizations need clear boundaries for their actions, security checks informed by the systems being changed, and an accountable owner for decisions to release.

Both of this week’s outages were caused by infrastructure changes, not cyberattacks. The common thread is that a change to one part of a complex system cascaded through interdependent components. As AI accelerates the pace of code changes flowing through those systems, the verification gap widens. Organizations need to govern what gets built, guide how it gets built, verify that it’s safe to deploy, and prove it — not just hope the change didn’t break something downstream.

That requires investment in the whole path from proposal to production, including reliable tests, security review, dependency and architecture context, controlled deployment, monitoring, and recovery. It also requires measures that assess more than output volume. Teams should know whether changes are safe, whether they create rework or incidents, and whether the organization can recover quickly when assumptions prove wrong.

AI can help developers and agents produce software at a pace that was difficult to sustain manually. The lasting gains will depend on whether the organizations using it can evaluate and govern those changes with equal discipline. Development speed matters; confidence in what’s being released matters just as much.

Manoj Srivastava, chief product and technology officer, Veracode

SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.

Manoj Srivastava

Manoj Srivastava is chief product and technology officer at Veracode.

Get daily email updates

SC Media's daily must-read of the most current and pressing daily news

By clicking the Subscribe button below, you agree to SC Media Terms of Use and Privacy Policy.

You can skip this ad in 5 seconds