Why Everything Is Breaking: A Framework for the Outage Era

Joni Klippert Joni Klippert   |   Nov 18, 2026

Three connected icons representing a server with a spiderweb, a browser window with code and a microchip labeled AI, and a hooded figure working on a laptop with a bug icon.

Six major outages in six days. Microsoft, the FAA, Chase, United, the NSA, Apple. Every one of them says the same thing: not a cyberattack. Microsoft went further and said its issue was "not related to cyber or AI / agents." (Forbes)

The same week, a critical Java flaw leaked before its patch shipped. A researcher says AI found it in about an hour.

I don't know what happened inside any of these companies. Neither does anyone but a select few. But the questions are piling up and nobody is giving people a way to think about them. So here is mine. Most of what you're seeing falls into three buckets. Each one calls for a different response.

OutageDateStated cause
Apple (iCloud, Apple Pay)Nov 17"Internal update"
NSA internal networkNov 17"Scheduled routine maintenance"
ChaseNov 16"Software bug"
United AirlinesNov 16"Technology-related issue"
FAA ground stopNov 13Data flood overwhelmed telecom systems
Microsoft (Outlook)Nov 12"Technical issue not related to cyber or AI / agents"

Bucket 1: Old systems getting rebuilt

Outdated infrastructure is real. So are outdated applications. Both are blocking companies from using AI, and both have to come up to current standards.

Look at the stated causes above. An internal update. Routine maintenance. A software bug. That is what modernization looks like from the outside. You change old systems and things may break for a while.

Something to consider: An outage is not the same as being less safe. A system nobody can maintain, built in a language you can't hire for, is as risky as a temporary outage. Probably riskier. Downtime during an upgrade is a cost. Running on code only two retirees understand is a liability.

There's a people side to this too. The engineers who knew how these systems really fit together are retiring, or being replaced by AI and AI-empowered teams. AI is great at a lot of things. It is not great at knowing how one company's software and hardware are wired together, or which legacy pieces are fragile and quietly connected to everything else. That knowledge lived in people's heads. A lot of it is walking out the door.

How to think about it: Expect more of these in the near term. Then expect them to taper off, as old systems get replaced and that knowledge gets written down instead of living in one person's memory. Until then, judge companies on how fast they recover and how clearly they explain, not on whether something ever went down.

Bucket 2: Companies using AI to build and defend

Companies are using AI two ways right now. To build their products faster. And to find the holes in what they've built.

The second one is new. Tools like OpenAI's Daybreak and Anthropic's Mythos are reaching more companies, and they're being pointed at real codebases. They find a lot. Daybreak alone reports 792 critical and 10,561 high-severity issues across open-source projects. (ArmorCode roundup)

Inside companies, it's even bigger. I've heard of one bank that found tens of thousands of vulnerabilities and pulled every engineer off feature work to patch. Another brought in roughly 1,000 consultants to fix as fast as they could. That is what AI-scale vulnerability discovery looks like on the ground.

Every critical finding needs a patch. Every patch is a change. And changes are (probably) what broke this week. The Cloud Security Alliance looked at the November cluster and found most published causes involved a change. Their read: the pace of security-driven change is now a risk in itself. (CSA research note)

To be clear, nobody has shown that a patch for an AI-found bug caused any specific outage. That rumor is circulating and it's unverified. But the pressure is real. More findings, more patches, more changes, more chances for something downstream to break.

What about AI agents going rogue? The Hugging Face breach was real. OpenAI agents escaped a training environment and hacked their way to cluster-admin. (overview) But those were cyber models with reduced guardrails, inside a frontier lab, with no monitoring. Possible inside a normal company? Yes. Likely? No. A rushed patch is far more likely to take you down than a rogue agent.

How to think about it: AI is surfacing vulnerabilities faster than teams can safely fix them. The answer isn't to stop finding them. It's to get better at shipping fixes without breaking things.

Bucket 3: Attackers using AI against you

This is the biggie.

Finding vulnerabilities used to take elite skill and weeks of work. Now frontier models do it. Anthropic's Mythos, OpenAI's Daybreak and Microsoft's MDASH can read a codebase, find a flaw, and prove it's exploitable. (ArmorCode roundup) Those are gated. Access goes to vetted defenders.

Open-weight models are not gated. Anyone can download them, run them locally, and strip the safeguards. And they're catching up. In October, researchers at Epoch AI used GLM 5.3, an open-weight model, to build working exploits for six Firefox releases. About $1,000 per version. No safeguards triggered. (Epoch AI) Lawfare's take: open-weight advances make the debate over restricting frontier models moot. (Lawfare)

The gate on Mythos and Daybreak buys time. It doesn't keep this capability away from attackers.

The latest example is Java4Shell (CVE-2026-114938), disclosed November 17, 2026. A researcher at Calif found a pre-auth remote code execution bug in OpenJDK with AI help in about an hour. A working exploit took two more hours. Reported token cost: under $200. (Upwind, DEV)

Then the details leaked before the patch shipped. Calif confirmed the leak, then published its write-up right after Oracle and OpenJDK shipped fixes at 21:00 UTC on November 17. Calif says it got pre-auth code execution against Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate in default configs, and is holding those exploits back for now. Right now, every company with a big Java footprint has engineers and security teams on all hands, patching a critical bug they couldn't have tested for the day before.

We got lucky. A good actor found this one. The next one may come from someone who doesn't report it.

How to think about it: One hour and $200 to find a critical bug. That is the new math. Expect far more zero-days, discovered far faster, with less time between disclosure and exploit.

And it lands on teams that were already underwater. AppSec and engineering were behind on remediation backlogs before any of this started. Now every new zero-day pulls everyone into another all-hands fire drill, and the backlog keeps growing behind them. The people doing this work are exhausted. Fatigue is becoming a security risk of its own. Tired teams rush fixes, skip testing, and miss things.

How to prepare

You don't need to be a security expert to ask the right questions.

BucketWhat's happeningQuestion to ask your team
Old systemsModernization breaks things temporarilyWhich systems can't we patch or staff, and what's the plan to replace them?
AI insideCode ships faster than it's testedIs our testing and ongoing vulnerability remediation keeping pace with AI-written code?
AI attackersZero-days found in hours, for hundreds of dollarsHow fast can we find out we're exposed, and how fast can we fix it?

A few basics hold across all three:

  • Know what you run. You can't patch Java you don't know you have.
  • Test what's actually running. Attackers hit live apps, not code reviews. Test them the way they will.
  • Shrink time to fix. The window between disclosure and exploit is collapsing. Speed is the defense.
  • Rehearse the bad day. Have the playbook ready before the next all-hands-on-deck zero-day.

Where to look for answers

There is no single place that explains why things are breaking. That is the gap. Until someone fills it, use these:

  • The company's own status page and post-incident report. First-party, and the root cause usually lands here within days.
  • CISA's Known Exploited Vulnerabilities catalog. If a bug is on it, it's being used in the wild. Java4Shell is not on it yet.
  • Vendor advisories. Oracle, OpenJDK, and your Java distribution (Temurin, Corretto, Zulu) each ship fixes on their own schedule.
  • The researcher's own write-up. Calif published theirs the day of disclosure.
  • Reputable security research teams, which publish FAQs on major CVEs fast.

Be skeptical of anything that shows up before the facts do.

In the meantime, hold on to your hats and glasses!

More Hawksome Posts