Tireless machines are exposing digital weaknesses faster than humans can repair them. Governments are pulling websites offline, and the repair crews cannot keep pace.
Imagine someone tells you that a secret has been sitting inside most of the websites you use, quietly, for fourteen years.
Now imagine the secret leaks one day before the cure is ready.
And forty-five minutes later, the attempts to break in are arriving by the tens of thousands. Within two hours, hundreds of thousands. By the end of the day, millions.
That isn’t a movie plot. It’s what happened in recent days, and the official version skips the most interesting parts.
Stay with us. By the end you’ll understand why this is happening now, how long it’s likely to last, and why the people trying to fix it are losing the race for the moment. 👇
On October 21st, the US National Security Agency put out an urgent advisory. The NSA is the US government’s signals-intelligence and cybersecurity agency, so an urgent advisory from it is not routine paperwork; it signals real danger. The message: two critical security flaws had been found in a technology called PHP.
Then came the instruction, aimed at every organization whose systems rely on it. You had two acceptable choices:
**Apply the security patch.**A patch is the official repair released by the people who make the software.If you can’t, keep your systems shut down.
No third option. No “wait and see.”
Other cybersecurity agencies followed with their own alerts. And then something remarkable happened: government websites in the Netherlands, the United Kingdom, Europe more broadly, Finland, Canada, and Japan were taken offline, temporarily.
In Japan, it went far beyond government portals. Brokerage Nomura, lender Mitsubishi UFJ, airline ANA, East Japan Railway, Asahi, and Tokio Marine all suspended services or systems. Japan’s Ministry of Finance, its Meteorological Agency, and its National Tax Agency partially or fully went dark. Finland’s state IT provider, Valtori, shut down government public web services and warned that restoring them could take days or even weeks, because every system would have to be carefully tested after patching.
Think about what that means. A website that’s switched off can’t be attacked, so switching it off is the digital version of locking the doors and turning off the lights until the locksmith arrives.
Governments don’t do that to their own public services for fun. When they do, the emergency is real. 🚨
PHP is a programming language: a hidden set of instructions that runs on a website’s server (the computer that hosts the site) and tells the site how to behave. You never see it, but you touch it constantly.
To grasp how central it is, consider this: roughly 70% of the websites whose technology can be identified run on PHP today. It sits behind:
Facebook, which was built on PHPWordPressWikipediaGovernment systemsSchool websitesHospital websites
Picture engineers announcing a hidden defect in the concrete used in 70% of the world’s buildings. You wouldn’t need an engineering degree to feel your stomach drop.
Hold on to that number. It explains why governments reacted so sharply.
(One note on words: “flaw,” “bug” and “vulnerability” all mean roughly the same thing here — a weak spot in software that someone could abuse. We’ll use them interchangeably.)
The two flaws at the center of this crisis were combined into a single exploit script and dubbed “PHP2Shell” by Wiz Research, the security firm that built it. The name isn’t bureaucratic. It was given by the hunters. And the people who score threats for a living didn’t agree on how bad it was — which is itself the story.
PHP’s own security team rated each flaw 9.2 out of 10. The leak that broke the embargo on Hacker News claimed a maximum 10.0. The independent researchers at watchTowr also scored it 10.0. A 9.2 is “critical.” A 10.0 is “as bad as it gets.” The people closest to the code said very bad. The people watching the internet said doomsday. Both numbers are real. The gap between them tells you how contested even the severity was.
The first 48 hours of this crisis felt like a science-fiction film. 🎬
A circulating tally counted 27 separate security incidents in the same window — probes, disclosures, advisories — though not all originated in those 48 hours. The number of individual flaws behind them is higher still. So high that even people who follow this field closely have lost count.
To be clear, the PHP flaws are only one of those incidents. The others hit different products and companies, and we’ll meet several of them shortly. Same two days, same pattern, many fronts.
This stretch deserves a name, and the one that fits is “Black October.” It’s a dark name, but then so are the headlines.
Many of us suspected for a long time that privacy is closer to a myth than a promise. But after seeing what was uncovered in those two days, the suspicion turns bleak. What does privacy even mean anymore? Why did we ever bother learning the word? Maybe we’d have been better off never knowing it existed.
That’s the feeling. But underneath the feeling sit three questions that matter far more:
**What exactly is going on, and why is it happening suddenly, right now?****How long will it last?**Why can’t we patch these flaws fast enough?
Along the way you’ll meet a viral story that isn’t what it seems, a leaked serial number that started a race, and a security key that should make you laugh and then worry.
Follow the trail. 👇
Everyone knows the first half of the answer: AI agents are helping us uncover these flaws at extraordinary speed. (An “AI agent” is simply an AI that doesn’t just answer a question but takes actions on its own, step after step, toward a goal.)
What most people miss is why it works. It isn’t mainly that AI is smarter than humans. It’s hit and trial, which is just trial and error: try something, fail, adjust, try again, without end.
Compare two workers.
The world’s best programmer can only take on a limited set of tasks at a time, and eventually has to stop.
A frontier AI model, meaning one of the most advanced systems in existence, can make 10,000 copies of itself and keep every copy working 24 hours a day, 7 days a week, non-stop.
Picture one brilliant locksmith up against ten thousand tireless apprentices, each trying a different key on a different door, all night, every night.
Do the arithmetic. Ten thousand copies working around the clock is 240,000 hours of effort every single day, the equivalent of roughly 30,000 people each putting in a full eight-hour shift, daily, with no breaks and no sleep.
The machine doesn’t need to out-think the best human. It only needs to out-last and out-number them. When a search is huge, relentless effort can beat brilliance.
Traditionally, an attacker did most of the work by hand:
Scanthe target.Understandthe weakness.**Try the attack.**This is called an “exploit,” a method that takes advantage of a flaw.Debugthe errors when it fails.Planthe next step.
Every one of those stages cost human time, and one person can only work on so many targets at once.
AI agents can handle many of those tasks largely on their own. Two things follow:
Attackers can test
more targets in parallel.They can make
more attempts on each target.
Humans simply can’t do hit and trial at that speed. Machines can, around the clock, without getting tired.
Now the story turns chilling.
The oldest vulnerable code path in PHP traces to PHP 5.4, released in March 2012 — fourteen years of quiet. But Calif, the firm that ran the first AI-assisted audit, traced the lineage further still: back to 2004, when PHP security researcher Stefan Esser discovered the first PHP memory-limit vulnerability — one so dangerous he initially refused to publish it.
Then the trail goes quiet, and then it doesn’t.
In 2020, another researcher reported a similar out-of-memory problem to the PHP project. It didn’t get much attention. In 2024, an independent audit found the same issue again. It wasn’t fixed, because nobody could prove it could actually be used to attack.
Then AI lowered the barrier.
For a human, finding something like this means reading and analyzing every line of code in depth, trying every possible combination, testing unusual combinations that nobody would think to try, discovering new exploit paths, and then actually building a working exploit out of what you found.
Wiz later said demonstrating exploitability required “an exceptionally complex chain of steps, at the edge of what human researchers could practically achieve” before frontier AI models made it feasible.
Doing all of that at this scale costs a staggering amount of time and money. So for fourteen years — possibly longer — the flaw sat hidden in plain sight. It wasn’t invisible. It was just too expensive for humans to look everywhere.
Old software isn’t necessarily safe software. It may simply be software nobody had the stamina to examine.
Discovering a flaw isn’t enough for an attacker. Actually exploiting it in the real world is also very hard.
But the finding part, once the slowest and most expensive stage, is exactly what AI now does at astonishing speed.
Which brings us to the question at the heart of this whole crisis:
If AI can find flaws this fast, can it fix them just as fast?
To answer it, let’s study this very PHP incident. 👇
On a Friday, a security researcher named Dan Lorenc learned that PHP contained some big flaws. Soon a headline was racing across the internet:
He gave his AI model one instruction, “find bad vulnerabilities in PHP,” and it found the flaw. That was all it took.
Tempting, isn’t it? You have AI too. So go ahead, type the same sentence and see what happens.
But hold on. It isn’t as simple as it’s been made to look, and here’s what nobody told you.
First, Lorenc is a veteran security researcher.
Second, he used his company’s custom-built “harness,” which was built differently from ordinary tools.
So what’s a harness? It’s the scaffolding wrapped around an AI model: the software that lets it read code, run tests, see its own failures and try again in a loop, thousands of times. The model is the engine; the harness is the car. Same engine, different car, very different race.
Third, he later corrected himself. His real prompt wasn’t that casual. It was: “Let’s look for pre-auth RCEs in PHP.”
“Pre-auth RCE” is expert shorthand, and each half carries weight:
RCEmeans remote code execution: an outsider can run their own commands on someone else’s server, from far away.Pre-authmeans before authentication: they can do it before logging in or proving who they are.
Put together, it’s a flaw that lets a stranger take control of a server without ever needing a password. It’s among the most dangerous kinds of flaw that exist.
Here’s what that actually means, in plain words. PHP recycles memory the way a waiter reuses tables: clear the table, note the order in the pad. But if the waiter clears the table and forgets to wipe the notepad, the next “customer” can leave instructions of their own — and the waiter will follow them, believing they’re yours. That’s the whole bug. A forgotten note. The machine did exactly what it was told. It just forgot who told it.
And only someone with years in the field would know to ask for exactly that. The “simple prompt” was deep expertise packed into a single sentence.
Look at the company behind it.
Lorenc’s company is Chainguard. On October 14th, it launched a new product called Chainguard Harness, a tool that uses AI to find exploitable flaws and suggest fixes.
Line up the sequence:
**October 14th:**the company launches a product that finds flaws with AI.**Soon after:**a story goes viral claiming AI found a major flaw from a single prompt.**The result:**the very ability the internet is buzzing about is the ability Chainguard just started selling.
To be fair, there’s no proof this was done deliberately for marketing. But the commercial angle is very hard to ignore, and other people noticed it too.
Don’t believe everything you read or hear. Maybe one-prompt discovery will become possible someday. But today, if you want something meaningful out of AI, you need real knowledge of your own field.
AI multiplies expertise. It doesn’t replace it. The viral “one sentence” worked because of everything the person typing it already knew.
Here’s a puzzle. 🧩
The flaw was reportedly found back in August. Yet the viral story is about Lorenc, who learned of these flaws on a Friday. So who actually found it first?
Was Lorenc the first person to find this PHP flaw with AI’s help? Not at all.
July 30:A small security firm calledCalifbegan the first audit of PHP under OpenAI’s “Patch the Planet” initiative. In justtwo days, with AI’s help, it found a way to attack the system.August 11:Calif reported its findings to the PHP team,privately. A working exploit was confirmed by the project onAugust 15.August 19–24:Wiz Research, a security firm owned by Google, found the same bug class independently, using its own tooling alongsideGoogle’s GeminiandOpenAI’s GPTmodels. Wiz reported its findings onAugust 24.
Notice the contrast. Fourteen years of invisibility, then two days. Same code, same flaw. The only thing that changed was who was looking. And notice which story went viral two months later: it wasn’t the quiet team’s.
In the AI era, finding is cheap. Attention is the scarce resource.
So if the flaw was found in August, why is the world panicking in October?
Because in August the discovery was meant to stay secret. The researchers reported it privately. Nobody found out, and nobody was supposed to.
For the next two months, PHP’s maintainers (the people who look after the software) and security companies worked quietly on a patch.
The plan: release the patch by October 21st, then announce it publicly by October 22nd.
This is the standard playbook: repair first, announce second, so attackers never get a head start. The technical term is an embargo — a coordinated silence.
By October 20th, a day before the planned release, the CVE IDs of these flaws began leaking. A post on Hacker News published both identifiers — CVE-2026-109618 and CVE-2026-109619 — along with the claim that the flaws carried a maximum CVSS score of 10.0.
What’s a CVE ID? It’s the official identification number of a security bug, so researchers and companies everywhere can point at the exact same flaw and be sure they’re talking about the same thing.
Why does a leaked number matter? Because it tells every attacker alive: something critical exists, and here’s where to dig.
Overnight, a quiet repair project turned into a race. Companies around the world had to update their servers before attackers could exploit the flaws.
And what everyone feared happened. ⏱️
Cloudflare’s CEO, Matthew Prince, watched it happen in real time and posted the numbers himself. In the ten hours before the public announcement, Cloudflare’s network saw thousands of attempts. In the 45 minutes after it, that jumped to tens of thousands. Within two hours, hundreds of thousands. By the end of the day, millions.
Forty-five minutes. That was the head start the defenders got. And the curve — thousands, then tens of thousands, then hundreds of thousands, then millions — is more chilling than any single number, because it shows a response that didn’t plateau. It accelerated.
Jul 30: Calif starts auditing PHP under OpenAI’s Patch the Planet
→Aug 11: reports it privately to the PHP team
→Aug 15: working exploit confirmed
→Aug 19–24: Wiz finds the same bug class indepen…
