Why Everything Broke at Once: AI Found the Flaws, and the Rush to Fix Them Caused the Tech Outages

In one week, Microsoft's email went down, the FAA grounded every flight in America, JPMorgan's systems stalled, Apple Pay failed, and the NSA's classified network went offline. Officials now say none of it was a cyberattack. The cause was the scramble to patch flaws that artificial intelligence models had just found — and the patches themselves knocked old systems over.
The week everything broke
The failures landed on consecutive weekdays. On Thursday, November 12, Microsoft 365 and Outlook went down most of the day. On Friday, November 13, the FAA ordered a nationwide ground stop — every departing flight halted for over four hours, the longest since September 11, 2001. On Monday, November 16, a software bug stalled JPMorgan Chase's systems for most of the day; because Chase sits inside so many payments, delays rippled to customers at Bank of America, Wells Fargo, U.S. Bank and Truist, and across the Zelle payment network. United Airlines had an outage too. On Tuesday, November 17, Apple services including Apple Pay went down after an update, and the NSA's classified network went offline.
We covered several of these, and the pattern was the same: a real failure, no evidence of a hack. The cyberattacks post made the point directly — an outage is not a hack. The Microsoft outage traced to a network capacity upgrade, the air travel shutdown to a faulty router. What those posts could not explain was why so many systems failed in the same few days. Now there is an official answer.
The flaws AI found
The answer starts with two flaws in the world's software, both AI-found.
The first is called PHP2Shell: two critical vulnerabilities in PHP, the programming language behind roughly 70% of the world's websites. Researchers at Wiz, Calif and OpenAI found them with frontier AI models, part of OpenAI's "Patch the Planet" effort. The flaws let an attacker run code on a web server without logging in, and reach back to a version of PHP from 2012. PHP shipped emergency fixes on October 21, and within five days the Cybersecurity and Infrastructure Security Agency (CISA) had listed both as actively exploited.
The second landed this week. On Tuesday, November 17, the security firm Calif disclosed Java4Shell, a critical flaw in OpenJDK — the open-source platform under Java, the language behind banking, hospitals, government and much of the Fortune 500. Calif researcher Quang Luong found it in an hour using OpenAI's Codex and built a working exploit in two more, for under $200. Wiz's AI system produced a proof-of-concept exploit within two hours and confirmed takeovers of Jira, Confluence and Jenkins. Oracle, which owns Java, shipped patches the same day, and CISA told organizations to patch immediately; as of the disclosure, no attacks using the flaw had been reported in the wild. A day later, a firm disclosed remote-code-execution flaws in NGINX, the web server in front of a large share of websites, also AI-found.
Java4Shell also shows who sets the clock. Oracle, which owns Java, scored the flaw 8.1 — "high," "difficult to exploit" — while Calif and Wiz scored it 9.2, "critical." Part of the gap is two scoring systems, but the effect is real: a 9.2 forces emergency patching, an 8.1 can wait. Oracle shipped its fix the same day anyway, outside its normal quarterly cycle. And because Java sits under banking, government and healthcare systems, when Oracle patches, thousands of companies must patch with it, on Oracle's timeline — though open-source alternatives like Corretto and Zulu, which shipped same-day fixes, check that power.
The pattern is the point. AI now finds critical vulnerabilities in foundational software — the languages and servers everything is built on — in hours, for the price of a nice dinner. And every find sets off the same scramble: patch now, before attackers build exploits for the flaw.
What officials told Congress
On Wednesday, November 18, the Director of National Intelligence, Jay Clayton, told the House Homeland Security Committee that the week's outages were "separate service outages associated with maintenance and technical changes. They are not a coordinated or state-sponsored attack." He tied the NSA outage to patching for the PHP flaws. The White House put it plainly: institutions are patching on a much larger scale, and "as they upgrade them, they're not coming back as expected in all cases." Bank chief executives and NSA officials told Congress privately that rushed fixes for AI-found flaws caused the outages, per three people familiar with the talks.
So the chain, as officials describe it: AI finds flaws faster than ever, forcing companies and agencies to patch urgently, all at once. The patches land on old, fragile systems never built to be changed at that pace — and some break. The outages were not attacks. They were the cost of fixing, in a hurry, what AI had just exposed.
The other side
That explanation is official, but not the whole picture. The FAA ground stop still traces to a single cause: a router at El Dorado, Arkansas, installed during a copper-to-fiber upgrade, that flooded other routers until the network gave up — a hardware story, not a patching story. The FAA's AI tool SMART, nationwide three days earlier, was cleared by officials on November 16; it runs on a separate data layer, unconnected to the failed network. Some security researchers dispute that the patch theory explains every failure. And one rival theory — that AI on old platforms repairs code faster than they can handle — rests on an unnamed expert, uncorroborated.
The honest summary is narrower than the headline. The patch rush is confirmed for the NSA outage and is the explanation officials are giving for the wave; the timing fits, with the PHP flaws disclosed in October and the failures clustering in mid-November. But each outage still had its own trigger, and at least one — the ground stop — was a router, not a patch.
What it means for you
Finding flaws got cheap. Fixing them didn't. An AI model can now surface critical software vulnerabilities in the code under the world's websites and banks in about an hour. Patching it across thousands of interconnected systems — carefully, without breaking anything — still takes teams days or weeks, and doing it in a panic is how you get a week like this. Expect more of it: as AI keeps finding flaws at this pace, emergency patch cycles that came a few times a year will come a few times a month, and each is a change to a system that can go wrong.
It is the same capability we covered in our DeepMind post, where AI found a faster way to break an old form of encryption. Pointed at code, AI finds the cracks. That is good news in the long run — the flaws were there, and now they are found before attackers find them. But in the short run, the finding is outpacing the fixing, and the gap between the two is where systems go down. For you, the upshot is simple: when your bank, airline or email goes offline this winter, it is more likely a rushed fix than an attack.
Sources & further reading
Checked November 19, 2026. Wiz Research, PHP2Shell disclosure, October 2026 (CVE-2026-109618 and CVE-2026-109619, both CVSS 9.2; found with frontier AI models by Wiz, Calif and OpenAI; PHP emergency fixes October 21, 2026). CISA Known Exploited Vulnerabilities catalog, October 26, 2026 (both PHP flaws added). Wall Street Journal, "AI Just Found a Bug That Puts Up to 70% of the World's Websites at Risk," October 2026. Calif Research, Java4Shell disclosure, November 17, 2026 (CVE-2026-114938; found in about an hour with OpenAI's Codex, working exploit in about two more hours, under $200; demonstrated against Jira, Confluence, Jenkins, Keycloak, JFrog Artifactory and PingFederate). Oracle Critical Patch Update, November 17, 2026 (Java fixes shipped same day). CISA advisory, November 17, 2026 (patch Java immediately; no in-the-wild exploitation reported as of disclosure). Wiz Atlas, Java4Shell proof of concept within two hours, November 2026. Reuters, "No coordination or foreign state behind recent US outages, Clayton says," November 18, 2026 — stacks921.com (DNI Jay Clayton, House Homeland Security Committee: "separate service outages associated with maintenance and technical changes"). The White House, "'No Indication of a Cyberattack': DNI Clayton and Secretary Rubio on Recent Tech Outages," November 18, 2026 — whitehouse.gov ("as they upgrade them, they're not coming back as expected"). Washington Sun, November 17, 2026 (bank CEOs and NSA officials privately told Congress rushed patches for AI-found flaws caused the outages; three sources). Associated Press, November 16, 2026 (FAA, DHS, DOT and Air Space Intelligence clear the SMART AI tool; separate data layer, no connection to the failed network). SiliconANGLE, "Recent spate of outages attributed to technical issues amid accelerated AI bug discovery," November 18, 2026 — siliconangle.com. Xint, NGINX2Root disclosure, November 18, 2026 (NGINX remote-code-execution flaws found using OpenAI models). The claim that rushed patches caused the outage wave is the explanation given by officials and reported from private congressional briefings; the FAA ground stop is separately traced to an L3Harris router, and some security researchers dispute the patch theory for every outage.
PHP, Java, OpenJDK and NGINX are trademarks of their respective owners, shown here for identification only. No endorsement is implied. The hero image is an AI-generated abstract, not a photograph.
