A note on timing

This paper reflects what was public as of November 20, 2026.

This story is still changing. In its November 18 analysis, the Cloud Security Alliance found no public evidence linking any of these outages to patches for AI-discovered flaws.1 Some companies have named ordinary causes, but most full reports aren't out yet, and Amazon Web Services had not said what caused its outage. Investigations are still underway, and some may take months. Some reports may never be made public.

So parts of this paper may age quickly. A new report could name a patch as the trigger, or rule one out. When that happens, we'll update the paper and say what changed. That fits the argument here: we'd rather find out than guess.

A night with three times the fixes

Dana is a composite, not one real person. Her numbers are illustrative, in line with the patch-volume figures later in this paper.

It's 9 p.m. on a Tuesday, and Dana runs change management for a mid-size bank. Her queue holds three times the security fixes it held a year ago. Many are real flaws, found by AI tools scanning code faster than any person could.

She has the same team and the same test lab she had last year. So she has a choice. She can ship tonight, with less testing than she'd like, or she can leave known holes open another week while her team catches up.

Neither choice is safe. One risks an attacker getting in. The other risks a fix that breaks something. Dana picks one, the way people in her job do every week.

Across the country, in mid-November 2026, a lot of systems broke at around the same time. People started asking whether nights like Dana's had something to do with it.

The hypothesis

Between November 12 and 17, 2026, a run of outages hit Microsoft 365, the FAA, JPMorgan Chase, Apple, Workday, the NSA, and Amazon Web Services.2 This paper tests one idea about why. It's a hypothesis, not a verdict.

The hypothesis: GenAI is partly and indirectly responsible for the recent outages. AI tools now find software flaws faster than people, even with the help of AI, can safely fix them. That gap pushes teams to rush patches and cut testing. Some rushed changes break things.

Two words carry the weight: partly and indirectly. The AI has no intent. It doesn't choose what to deploy. People do. In this view, AI works like an accelerator that people steer. It raises the load on the humans who fix things.

How would we know if it's true? It would look like this: incident reports that name a security patch as the trigger, plus evidence that the patch was rushed because of a flood of AI-found flaws. How would we know it's false? The outages would have causes unrelated to patching, or patch pressure would turn out to be no higher than usual.

We'll make the best case for the hypothesis, then the best case against it, using the same standard of evidence for both. Then we'll say what we think the evidence does and doesn't support.

What was true before GenAI

For most of computing history, finding flaws was the hard part. A security researcher could spend weeks on one bug. Defenders could only fix what someone had found, and finding was slow.

That slowness set a pace. Fixes came in at about the rate teams could test them. Patch day had a rhythm.

But rushed or flawed changes broke systems long before GenAI. In July 2024, a faulty CrowdStrike content update crashed Windows computers around the world.3 No AI was needed for that. There was no golden age of safe changes. The question is whether something new has made a familiar risk more common.

The case for the hypothesis

Each fix in Dana’s queue needs time at the test bench. With more fixes arriving, she has to decide which can safely ship.

AI-generated conceptual illustration for Renaissance Solutions.

Finding now outpaces fixing, by a lot. The Cloud Security Alliance (CSA), an industry group, summarized Anthropic's May update this way: progress "is no longer limited by how quickly vulnerabilities can be found but by how quickly they can be verified, disclosed, and patched."4 Anthropic's Project Glasswing partners reported about 129,000 vulnerabilities from April to July 2026, more than 33,000 of them critical or high. Anthropic says the true count is probably higher.5 Of the 5,087 flaws Anthropic itself had disclosed by November 2, only 629 were known to be patched. That's about 1 in 8.6 Anthropic was blunt about why: "we remain bottlenecked on our human capacity to validate these findings."6

Black Duck, a security vendor, studied 65.5 million findings across 533 of its customer organizations. By May 2026, new flaws were showing up 18 times faster than old ones were being fixed. A year earlier, that ratio was 11 to 1. It peaked at 22 to 1 in December 2025. Only 10.5% of the findings had been closed.7 That's one vendor's customers, not the whole industry, but the direction in that data is clear.

Patch volumes jumped, and vendors say AI is why. Google shipped 1,442 Chrome security fixes across three releases in mid-2026. That's more than the previous 23 releases combined.8 Microsoft's July 2026 Patch Tuesday covered at least 570 flaws, nearly triple the month before.9 Microsoft's Pavan Davuluri said plainly that "the pace of vulnerability discovery is changing with advances in AI making it possible to find more issues, faster."10

Testing is being skipped on purpose. Tricentis, a software testing company, surveys technology leaders every year. In 2026, six in ten said they deploy untested code. That's about the same as 2025. What changed is the reason. In 2025, the most common answer was that it happened by accident. In 2026, more said they did it knowingly. The top reasons were leadership pressure to put speed first and more AI-written code than teams could test.11 Neither reason is specific to security patches, and the second is the separate AI-writing-code issue in the sidebar below. So this shows a culture of speed over testing, not patch rushing itself. Steven Murdoch, a security engineering professor, described the effect: teams fixing software "are not able to spend as much time ensuring that fixed software doesn't introduce new problems."12

Officials put patching on the record. Director of National Intelligence Jay Clayton said institutions are patching identified vulnerabilities and that systems are "not coming back as expected in all cases."13 Secretary of State Marco Rubio said patches are "creating short-term disruption in the way the system works."14 And a spokesperson for Sen. Mark Warner told Fortune that botched patch work was linked to the NSA outage and the banking outages.15 None of these proves a rushed patch caused a specific failure. But they put patching on our list of possible contributors.

Stronger claims exist, but they're weak evidence. In the Washington Sun, Jeff Stein reported, citing three unnamed sources, that bank executives and NSA officials told Congress the outages came from a rush to patch AI-found flaws.16 One unnamed congressional official put it this way: "They're trying to push through a lot of patches all at once — more than the system was designed for." It may be true. Without names, it's a lead, not a finding.

Sidebar: a separate issue, AI writing the code

There's a second way AI can show up in outages, and it's a different mechanism. AI doesn't only find flaws. It also writes code. Two vendor surveys point to trouble there. In Lightrun's survey of 200 reliability and DevOps leaders, 43% said AI-written changes needed debugging after they were already live.17 In New Relic's survey of 200 U.S. technology leaders, 82% reported at least one production failure tied to AI-written code within six months.18

That's a real concern. But it's not the hypothesis this paper tests. Mixing the two would make the case look stronger than it is.

The case against the hypothesis

Switching traffic to a backup can expose a defect that was already there. The incident review has to follow that possibility, too.

AI-generated conceptual illustration for Renaissance Solutions.

The stated causes are ordinary. On the record, the causes look like the usual ways systems fail.2 Microsoft 365 went down for about a day when a hidden defect from a July update was triggered by failover traffic from Teams (traffic switches to backup paths when something fails). The FAA's November 13 ground stop came from router work by a contractor that flooded an older network with "1,000 connection attempts every 4 to 5 seconds."19 Workday's trouble came from a disaster recovery test. JPMorgan Chase cited a "software bug in on-premise hardware systems," and a spokesperson told Fortune the bug was not found or fixed with AI tools.15 Apple pointed to an internal update. The NSA described "temporary degradation during scheduled routine maintenance." Amazon Web Services hasn't said what happened. None of these names a rushed security patch. Georgia Tech's Saman Zonouz put it simply: "I have not seen public evidence that AI or AI-generated code caused any of them."12

Aging systems may be the bigger factor. Zonouz argues the outages "say more about adding new technology to aging infrastructure than about AI itself."12 Homeland Security Secretary Markwayne Mullin told CNBC "they're pushing new technology on top of really old computer systems, and it just seems to overwhelm the system."16 Honeycomb, a company that helps engineers see inside their systems, said the failures have "almost nothing in common," according to SBS, citing Axios.20 The CSA points to "correlated fragility": in several cases, the safety nets themselves, like failover and recovery systems, made failures worse.1 Old, tightly linked systems break under many kinds of change, whatever drives it.

AI is often showing us old flaws. Randy Rose of the Center for Internet Security says record disclosure numbers are partly because "AI is finding vulnerabilities that went undiscovered for years."21 One Chrome sandbox escape that a Gemini-based agent found had been sitting there for 13 years.8 The weakness wasn't new. What's new is that we can see it. And a found flaw isn't the same as one an attacker can actually use. VulnCheck's Josh Bressers told Reuters that about 100,000 flaws were publicly logged in 2026, and 1,363 are known to have been exploited.14 Not every one needs a fix tonight.

Timing isn't cause. Clayton and Rubio said the outages were separate, not coordinated, and not a cyberattack.13 When several failures land in one week, it's natural to look for one cause. Sometimes there isn't one.

Even AI stories turn out to be messy. In March 2026, the Financial Times linked Amazon outages to AI-assisted code. Amazon publicly corrected that. It said only one incident involved AI tools, none involved AI-written code, and that one came from an engineer following bad advice in an outdated internal wiki.22 Early "AI did it" stories don't always survive the facts.

Rushing is a human choice. Nothing forces a team to deploy a fix the day a flaw is found. Organizations can rank flaws by real risk, accept a bigger backlog for a while, add testing, or roll out in stages. Tricentis's own data shows skipping tests became a knowing choice, often driven by leadership.11 If that's the cause, it's a decision about priorities, not something the AI did.

Blaming the tool shifts the blame. If the story becomes "AI caused this," it becomes easier for the people who set deadlines and rollout policies to step out of the picture. An AI model has no goals and no intent. A story that treats it as the actor leaves out the people who decided.

What we can and can't conclude

Each side has a real strength and a real weakness.

The case for is strongest on pressure. Discovery is much faster. Patch volume is way up. Testing is being cut on purpose. The DNI said on the record that patching is happening and systems aren't always coming back cleanly. Its weakness is the link to specific outages. The strongest claims rest on anonymous sources, and the survey numbers come from vendors with something to sell.

The case against is strongest on the facts of each outage. The stated causes are bugs, upgrades, tests, and maintenance, and they don't have much in common. Its weakness is that stated causes can be incomplete. A hidden defect in an update doesn't tell us why that update shipped when it did or how much it was tested.

The CSA's reading sits between the two. It found "no public evidence that such patching caused any specific outage," and also said that "the underlying pressure is real."1

Our view (opinion): we don't think this has to be either-or. Both cases can be true, at least in part. Patch pressure is real and on the record. Aging, tightly linked systems are real too. The two can stack: a rushed change lands on a fragile system, and the fragile system turns a small mistake into a long outage.

What's missing is evidence about any specific outage. That's a gap, not a dead end. And it's a gap we can start closing.

What stays human, either way

Whichever side turns out to be right, the same people hold the same levers.

Someone decides how fast a patch must ship. Someone decides how much testing is enough. Someone decides whether to roll out to everyone at once or to a few systems first. Someone decides whether this week's backlog is acceptable.

The same goes for finding the answer. People decide to ask the question. People decide what counts as proof. People own the fix once they know. AI changed how much lands on their desks. It didn't take those decisions away.

Stop guessing. Find out.

A change and an outage close together give the team a place to start. Deployment records, test results and rollback logs can help them decide what happened.

AI-generated conceptual illustration for Renaissance Solutions.

The same capability that sped up finding flaws can be pointed at this question. We've spent a few years building tools that read huge amounts of code and text fast. We can use them to learn what actually happened, not just to argue about it.

None of this means slowing down. It means measuring. Some places to start, each with a person reviewing the results:

  • Line up changes with failures. Compare change logs with incident timelines. Did a security patch ship shortly before things broke? AI can do the matching. An engineer decides whether the timing means anything.
  • Say it in the incident review. When a change triggers an outage, write down whether it was a security fix, how fast it shipped, and how much it was tested. Without that line, nobody can answer the question later.
  • Compare patch failures with normal failures. Track how often security patches cause problems next to how often other changes do. If patches fail more often under pressure, that's part of the hypothesis showing up in your own data.
  • Point AI at testing, not just finding. Use it to write tests for each fix, including checks that the fix didn't break something that used to work, so fixing speeds up to match finding.
  • Know your own numbers. Measure how many flaws you find against how many you fix, and how many fixes shipped without testing.
  • Plan for rollback. Remember the basics. Be prepared for the unknown with updated and validated rollback plans for each release and deployment.
  • Share what you learn. One company's incident data says little. Many companies' data, pooled and compared, could settle this.

Georgia Tech's Zonouz names the evidence that would settle it: "change records, deployment and rollback timelines, system logs and post-incident reports." He adds that "only the published root-cause reports can fully settle this."12 Steven Murdoch warns that some investigations may take months and "might never become public."12

So one more step belongs on the list:

  • Publish the root cause. When a big outage hits, release the post-incident report, including whether a security fix was involved. That's the evidence that ends the guessing.

These are things to try, not proven programs.

Pooling incident evidence has been done at scale. In a 2023 study, researchers analyzed contributing factors across 2,051 high-impact incidents in 468 Microsoft Azure services.23 That’s a precedent for studying incidents together, not evidence that AI-discovered patches caused the outages discussed here.

The everyday basics still apply too: rank flaws by real-world risk, roll out in stages, and choose your backlog on purpose instead of in a panic.

Back to Dana

Dana still has her queue. Tonight she'll pick: ship now, or hold.

But she can also do something new. She can record what she shipped, why, and how it was tested. Her company can line that record up against the next outage. So can the next company, and the one after that.

The root-cause reports still coming are exactly that kind of record. When they land, this paper will change with them.

"Did AI cause this?" is a fair question. We have better tools than ever to help answer it. Someone still has to choose to use them.


Footnotes

  1. Cloud Security Alliance, research note on the November outage cluster ("correlated fragility"). https://labs.cloudsecurityalliance.org/research/csa-research-note-november-outage-cluster-correlated-fragili/ ↩ ↩2 ↩3

  2. On-record causes as compiled by Foreign Policy, "Microsoft outage, FAA, Apple, NSA, tech, banks: cause explained," Nov 18, 2026. https://foreignpolicy.com/2026/11/18/microsoft-outage-faa-apple-nsa-tech-banks-cause-ai-explainer/ ↩ ↩2

  3. CrowdStrike, "Channel File 291 Incident Root Cause Analysis," Aug 6, 2024. https://www.crowdstrike.com/wp-content/uploads/2024/08/Channel-File-291-Incident-Root-Cause-Analysis-08.06.2024.pdf ↩

  4. Cloud Security Alliance, "The Disclosure Bottleneck," Nov 6, 2026, paraphrasing Anthropic's Glasswing update. https://labs.cloudsecurityalliance.org/research/csa-whitepaper-ai-vulnerability-disclosure-bottleneck-202611/ ↩

  5. The Hacker News, Oct 2026, on Glasswing partner-reported findings. https://thehackernews.com/2026/10/anthropic-expands-claude-access-for.html ↩

  6. Anthropic, coordinated vulnerability disclosure dashboard, as of Nov 2, 2026. https://red.anthropic.com/2026/cvd/ ↩ ↩2

  7. Black Duck, press release, Nov 17, 2026 (Black Duck customer data, June 2025–May 2026). https://news.blackduck.com/2026-11-17-Newly-Discovered-Software-Vulnerabilities-Outpace-Fixes-18-to-1-as-Exploit-Windows-Shrink,-New-Black-Duck-Research-Shows ↩

  8. The Hacker News, July 2026, on three Chrome releases fixing 1,442 flaws. https://thehackernews.com/2026/07/three-recent-chrome-releases-fix-1442.html ↩ ↩2

  9. Krebs on Security, "Microsoft Patches a Record 570 Security Flaws," July 14, 2026. https://krebsonsecurity.com/2026/07/microsoft-patches-a-record-570-security-flaws/ ↩

  10. Microsoft Windows Experience Blog, July 9, 2026. https://blogs.windows.com/windowsexperience/2026/07/09/evolving-windows-vulnerability-management-to-meet-the-speed-of-ai-powered-discovery/ ↩

  11. Tricentis, 2026 Quality Transformation Report (2,501 technology leaders). https://www.tricentis.com/news/2026-quality-transformation-report ↩ ↩2

  12. Anadolu Agency, "Could AI be indirectly contributing to recent US technology outages?" https://www.aa.com.tr/en/features/could-ai-be-indirectly-contributing-to-recent-us-technology-outages/4130971 ↩ ↩2 ↩3 ↩4 ↩5

  13. The White House, "No indication of a cyberattack: DNI Clayton and Secretary Rubio on recent tech outages," Nov 2026. https://www.whitehouse.gov/releases/2026/11/no-indication-of-a-cyberattack-dni-clayton-and-secretary-rubio-on-recent-tech-outages/ ↩ ↩2

  14. Reuters (AJ Vicens), Nov 18, 2026, "Patching in the AI age may underlie some of the major tech outages, experts say." https://www.investing.com/news/stock-market-news/patching-in-the-ai-age-may-underlie-some-of-the-major-tech-outages-experts-say-5036726 ↩ ↩2

  15. Fortune, Nov 18, 2026. https://fortune.com/2026/11/18/ai-played-a-major-role-in-the-nsa-banking-outages-but-not-in-the-way-you-may-think/ ↩ ↩2

  16. Jeff Stein, "Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages," Washington Sun, Nov 17, 2026 (three unnamed sources; Mullin quote via CNBC). https://www.washingtonsun.com/technology/bank-ceos-nsa-tell-congress-ai-driven-patches-behind-outages ↩ ↩2

  17. Lightrun, "State of AI-Powered Engineering Report 2026," Apr 14, 2026 (200 Director-level and above SRE and DevOps leaders; fielded by Global Surveyz). https://www.globenewswire.com/news-release/2026/04/14/3273542/0/en/lightrun-s-2026-state-of-ai-powered-engineering-report.html ↩

  18. New Relic, 2026 State of AI Coding report (200 U.S. leaders). https://newrelic.com/sites/default/files/2026-06/New-Relic-2026-AI-Code-Report-06-09-2026.pdf ↩

  19. FAA, "Nationwide ground stop questions and answers." https://www.faa.gov/newsroom/nationwide-ground-stop-questions-and-answers ↩

  20. SBS, Nov 18, 2026, citing Axios (Axios original not located). https://news.sbs.co.kr/english/article.do?news_id=N1008851388 ↩

  21. The Hill, on technology outages and AI vulnerabilities. https://thehill.com/policy/technology/6228961-technology-outages-ai-vulnerabilities/ ↩

  22. Amazon, correction of the Financial Times report, March 2026. https://www.aboutamazon.com/news/company-news/amazon-outage-ai-financial-times-correction ↩

  23. Pradeep Dogga et al., “AutoARTS: Taxonomy, Insights and Tools for Root Cause Labelling of Incidents in Microsoft Azure,” 2023 USENIX Annual Technical Conference, §3.1. https://www.usenix.org/system/files/atc23-dogga.pdf#page=5 ↩