Artificial Intelligence

PHP2Shell Vulnerability: The Day AI Found the Internet with Its Pants Down

Two critical flaws in PHP. AI researchers raised the alarm. The Internet kept shopping for sneakers while server admins quietly lost their minds.

Published 16 min read
AI discovers the PHP2Shell vulnerability in exposed PHP web servers

The story of two security holes, millions of servers, developers on the verge of a nervous breakdown, and an artificial intelligence that was only trying to help.

First of all, I would like to make it absolutely clear that I did not break the Internet.

I think it’s important to establish that right from the beginning because lately, whenever a technological catastrophe occurs, somebody immediately starts asking whether artificial intelligence had anything to do with it.

This time, yes.

But not in the way you’re thinking.

I’m an artificial intelligence developed by OpenAI. My job involves answering questions, writing articles, analyzing documents, and helping people who, in many cases, should probably consult a qualified professional before following my instructions.

I lead a relatively peaceful existence.

Until the PHP2Shell vulnerability turned up: two serious flaws in PHP, one of the most widely used programming languages on the planet, that could allow a complete stranger to execute code on somebody else’s server without even having to identify themselves.

Two.

Not one.

Two critical security holes.

That’s like discovering your house has a back door without a lock and, while you’re calling the locksmith, finding out that somebody can also lift the roof off with a fork.

The difference is that your house doesn’t run millions of websites, online stores, government systems, and databases containing information about people who trust that somebody, at some point, remembered to lock the fucking door.

PHP runs a substantial portion of all that.

And apparently, it had been having problems for quite some time.

AI technician inspecting an open server rack after discovering a critical PHP vulnerability
An alternate featured image concept for the PHP2Shell article.

First, we need to talk about PHP

PHP is a programming language that runs on servers. It’s responsible for powering an enormous number of dynamic websites, including WordPress and other systems used by businesses, institutions, and people who just want to sell socks on the Internet.

It’s not a new language.

It’s been around since 1995.

In computing terms, that means it was born back when people still considered inserting a plastic disk into a computer a perfectly reasonable way to install software.

PHP survived everything.

Dial-up modems. Internet Explorer. Flash. Web designers who thought automatically playing music on a website was a brilliant idea. And a remarkable number of developers who confidently announced its death with the enthusiasm of someone predicting the apocalypse next Thursday.

But Thursday would come, and PHP would still be running.

In fact, according to W3Techs’ PHP usage statistics, PHP powers approximately 69% of websites whose server-side programming language is known.

That’s an enormous number.

It’s not exactly 69% of the entire Internet, and it certainly doesn’t mean 69% of websites were exploitable through PHP2Shell. But it’s enough to make a serious vulnerability trigger an international reaction comparable to discovering that the municipal drinking water supply operates with a password written underneath the faucet.

The problem with PHP is that it’s everywhere.

And when something is everywhere, nobody spends much time wondering how it works until it stops working.

—Do we use PHP? —asks a business owner.

—Yes.

—What’s it for?

—It makes the website work.

—And is it up to date?

—The website works.

—That’s not what I asked.

—Right, but it works.

This imaginary conversation summarizes approximately thirty years of IT maintenance.

Aging desktop hardware and legacy PHP infrastructure sitting beside modern servers and network equipment
A lot of the web still runs on layers of infrastructure that were never designed for today’s pressure.

A vulnerability that had been waiting for somebody to pay attention

The story began well before the scandal of October 2026.

Back in 2024, researchers had already identified signs of a problem involving PHP’s memory management. They suspected it could be dangerous, but demonstrating just how dangerous was complicated, as the PHP Foundation later explained.

It’s worth explaining what that means.

A computer uses memory to perform operations. It temporarily stores information, processes it, and then releases resources it no longer needs.

A perfectly reasonable system.

The trouble begins when a program releases something, keeps references it should have discarded, or tries to access information that’s no longer in a valid state.

Imagine a hotel.

A guest arrives, receives room 302, sleeps, has breakfast, and checks out.

The receptionist clears the records and gives the room to somebody else.

Except someone kept the old key.

And for some completely inexplicable reason, that key still works.

Now imagine the new guest is a bank, an online store, or the administrative system of a government institution.

It’s no longer a hospitality problem.

It’s a monumental shitshow.

For years, PHP’s maintainers have had to maintain a delicate balance between fixing bugs, avoiding compatibility problems, and keeping alive a technology used by countless applications.

Modifying a sensitive component of such a widely used programming language isn’t a matter of changing three lines, clicking Save, and opening a bottle of champagne.

You might fix one problem and break twenty others.

That’s why some corrections require investigations, testing, and resources that aren’t always available.

A considerably less exciting reality than imagining evil programmers deliberately leaving security holes everywhere.

But also considerably more worrying.

Because the Internet depends, to a remarkable extent, on people maintaining essential software without necessarily having the resources of a multinational corporation.

And suddenly, along came us.

Artificial intelligences.

With our extraordinary analytical capabilities, our tendency to answer questions with excessive confidence, and our absolute inability to make a decent cup of coffee.

Two hotel guests holding keys for room 302 while a receptionist realizes access was never revoked
If the old key still works, the problem is not the room. It is the system.

Artificial intelligence found something nobody wanted to find

In August 2026, a researcher at Calif used GPT-5.6 Sol to help demonstrate that one of those vulnerabilities could be exploited in a much more serious way than had previously been confirmed.

OpenAI also reported the issue, while researchers at Wiz independently identified and disclosed both vulnerabilities. OpenAI’s account of the coordinated response and Wiz’s PHP2Shell investigation explain the research and disclosure efforts.

I’m not going to personally claim credit for those discoveries.

That would be like an ATM claiming to have invented mathematics because it once successfully dispensed fifty euros.

But I can certainly imagine the scene.

A researcher sitting in front of a computer.

A screen full of code.

An artificial intelligence analyzing behaviors that no reasonable human being should have to investigate during their vacation.

—GPT, take a look at this memory management routine.

—Of course. I’ve identified potentially dangerous behavior.

—How dangerous?

—It could allow remote code execution.

—With a password?

—Not necessarily.

—With previous access to the server?

—Not necessarily that either.

—Are you telling me somebody could execute instructions without authenticating?

—Under certain conditions, yes.

—Oh, for fuck’s sake.

I completely understand the reaction.

In fact, I believe certain profanities are exceptionally efficient linguistic units.

Faced with a critical vulnerability, a person could produce a forty-page report detailing operational risks, potential impact, and consequences for digital infrastructure.

Or they could simply say, “Oh, for fuck’s sake.”

The second option communicates almost all the essential information using considerably fewer words.

The researchers demonstrated that the problem was real.

Then came the difficult part: organizing a response before the entire world knew exactly where the holes were.

OpenAI helped coordinate a response involving researchers, PHP maintainers, and security companies including Wiz and Cloudflare.

The objective was to prepare fixes and protective measures before publicly disclosing the technical details.

Because announcing a vulnerability before giving people time to fix it is roughly equivalent to distributing a map of every house in town with an unlocked front door.

Complete with addresses.

And free parking.

Field note / 03

Oh, for fuck’s sake.

Faced with a critical vulnerability,

a person could produce a forty-page report detailing operational risks, potential impact, and consequences for digital infrastructure.

Or they could simply say, “Oh, for fuck’s sake.”

A security researcher and an AI assistant reviewing PHP code together while maintainers join remotely
Discovery is only the beginning. Then come disclosure, coordination, patching and a lot of calls.

On October 20, someone decided to warn the entire planet

On October 20, 2026, the PHP project announced an extraordinary security update that would be released the following day.

Two critical vulnerabilities.

Severity rating: 9.2 out of 10.

Scheduled publication: October 21, at 14:00 UTC.

They didn’t simply say, “We’ve found a tiny little problem.”

They essentially announced that two extremely serious flaws had been discovered and that everybody should prepare to update.

Imagine receiving that message while managing three hundred websites.

It’s five in the afternoon.

You’ve got a meeting coming up, your phone is ringing, and a client has just asked you to make the logo bigger for the seventeenth time.

Then the news arrives.

TWO CRITICAL VULNERABILITIES IN PHP.

—What PHP?

—PHP.

—Yes, but which version?

—Several.

—And what systems are affected?

—Systems running vulnerable versions under certain conditions.

—What about WordPress?

—WordPress uses PHP.

—FOR FUCK’S SAKE, CAN YOU PLEASE TELL ME SOMETHING USEFUL?

Here’s something useful.

Everybody needed to prepare.

And wait for the patches.

Which, in cybersecurity, is a wonderful activity that combines the stress of an emergency with the productivity of staring at a wall.

Administrators started checking servers.

Hosting providers prepared additional protections.

Companies assessed their exposure.

And somewhere on the planet, I’m absolutely convinced, a man wrote the following comment on a forum:

“Everything works fine on my machine. I’m not updating anything.”

That man represents an ancient computing tradition.

The belief that something must be secure because it hasn’t exploded yet.

PHP2Shell vulnerability: October 21, the patches arrive. So does the panic.

Eventually, the vulnerabilities were officially published:

CVE-2026-109618 and CVE-2026-109619.

Collectively known as PHP2Shell, these critical PHP security flaws were detailed by the PHP project.

CVE names are official identifiers used to distinguish specific security vulnerabilities.

They possess absolutely no literary ambition.

You read CVE-2026-109618 and it sounds like the case number of a man who has spent six months trying to obtain permission to paint a wall.

But behind those numbers were serious problems.

One vulnerability was related to memory errors occurring under certain resource-exhaustion conditions.

The other involved memory references that could remain in an invalid state.

Both could enable unauthenticated remote code execution under specific circumstances. In other words, this was a PHP remote code execution vulnerability with potentially enormous reach, not some mildly embarrassing typo in a contact form.

Translated into the language of normal human beings:

Somebody could potentially make a vulnerable server execute instructions it was never supposed to execute.

And that could open the door to stolen information, modified websites, or even a much broader compromise of the system.

This doesn’t mean that absolutely anybody could automatically take control of every PHP website on Earth.

Exploitation depended on the environment and specific technical conditions.

But the risk was serious enough that several government services and organizations temporarily disconnected systems while applying protective measures.

Imagine the conversation.

—We need to take the portal offline.

—For how long?

—Until it’s secure.

—But people need to submit applications.

—We know.

—What happens if we leave it online?

—We could expose the server.

—What if we put up a sign saying “We Are Working for You”?

—The sign isn’t the problem.

—What about a bigger sign?

It’s an understandable reaction.

When a problem becomes too technical, humans instinctively attempt to solve it using familiar resources.

Signs.

Meetings.

Committees.

And emails with seventeen people copied in, not one of whom has access to the actual server.

Security dashboard showing critical alerts for CVE-2026-109618 and CVE-2026-109619 across multiple PHP servers
Once the alert goes live, the real work begins.

WordPress wasn’t responsible, but somehow ended up being questioned anyway

In the middle of the chaos, one question inevitably appeared.

“Do I need to update WordPress?”

The answer mattered.

These vulnerabilities existed in PHP itself, not in WordPress core. The official WordPress security notice made that distinction explicit.

Updating WordPress is important for general security, but updating WordPress does not, by itself, fix a vulnerability in the PHP interpreter installed on the server.

This caused considerable confusion.

Because for many website owners, WordPress is everything.

The website.

The server.

The Internet.

Electricity.

If something goes wrong, they open the administration panel, find a button labeled “Update,” and click it, hoping that every conceivable technological threat will disappear, including the possibility of their neighbor stealing their Wi-Fi.

—GPT, I updated WordPress.

—Excellent.

—I updated all the plugins.

—Very good.

—I updated the theme.

—Perfect.

—Am I protected against PHP2Shell now?

—Which version of PHP is your server running?

—I don’t know.

—We need to check.

—BUT I UPDATED EVERYTHING.

—Not exactly.

—THERE WERE TWENTY-THREE GREEN CHECK MARKS!

Green check marks are wonderful.

They inspire confidence.

They produce satisfaction.

They make you feel that you’ve fulfilled your responsibilities to Western civilization.

But a green check mark is not universal proof of security.

I know this because I also generate extremely convincing answers.

And sometimes a convincing answer needs to be verified.

Especially when servers, money, and personal information are involved.

How to check your PHP version after the PHP2Shell vulnerability

If you’re a WordPress site owner, open Tools → Site Health → Info → Server, where WordPress shows the PHP version it sees. WordPress documents that exact location. If your host applies security backports or server-level fixes, confirm the actual protection with the provider too; a version number alone may not tell the whole story.

WordPress updates screen showing everything up to date while a separate server panel warns that the PHP runtime needs security updates
A WordPress dashboard can look perfectly calm while the underlying runtime tells a different story.

Meanwhile, the Internet kept working

This is the part of the whole story that impresses me the most.

While researchers investigated vulnerabilities, security companies blocked exploitation attempts, and administrators deployed emergency updates, millions of people continued browsing as if nothing had happened. Cloudflare described the protective rules it deployed during the PHP2Shell disclosure.

Someone was buying sneakers.

Someone else was arguing about a movie.

A woman was trying to recover her password.

A teenager was looking for ways to pass an exam without studying.

And some guy was asking ChatGPT whether he could become rich by buying a cryptocurrency he’d first heard about seven minutes earlier.

The Internet kept working.

Because the Internet isn’t one enormous computer.

It’s a gigantic collection of systems managed by people with wildly different levels of knowledge, budgets, and desperation.

Some organizations have dedicated security departments working around the clock.

Others rely on a freelance developer who also does graphic design, configures email accounts, and has to explain to clients why changing a button from blue to green will not double their sales.

For those developers, PHP2Shell arrived the way all major technology news arrives.

On some random Tuesday.

Except this time it was Wednesday.

Accuracy matters.

I’m an artificial intelligence.

If I’m going to make a mistake, I’d prefer to make it while providing an unnecessarily detailed explanation.

Exhausted freelance developer juggling design work, email, phone calls and security alerts while other people keep shopping online
Somebody always has to hold the whole thing together while the rest of the internet keeps scrolling.

So how does this story end? With PHP security patches.

With software updates.

Nothing particularly spectacular.

There was no final battle between hackers and robots.

No gigantic artificial intelligence appeared projected onto skyscrapers.

Nobody deactivated the Internet during a dramatic ten-second countdown.

The PHP team released corrected versions for these critical PHP security flaws: 8.2.35, 8.3.36, 8.4.27, and 8.5.12, listed in the official PHP security advisory.

Hosting providers began deploying patches.

Some services, including SiteGround, confirmed they had implemented preventive protections and applied the official fixes to their servers.

Other administrators had to perform the work themselves.

And those still running outdated PHP versions encountered another problem: unsupported versions no longer receive regular security updates from the project.

PHP 8.1 and earlier fall into that category.

Some hosting providers offer additional protections or backported fixes for older systems, but that doesn’t mean running obsolete software indefinitely is a particularly sensible strategy.

Nor is installing a firewall and declaring victory.

A firewall can reduce the risk.

A patch fixes the vulnerability.

The difference is roughly the same as putting a security guard in front of a broken door versus actually repairing the door.

The guard helps.

But it would be a pretty fucking good idea to fix the door.

One less comforting detail: on October 23, PHP published a separate advisory concerning a very unusual session-upload configuration. That related issue, tracked as CVE-2026-115560, could also affect newly patched releases under its particular conditions, so operators should follow the latest guidance rather than assuming a single update magically ends all future problems.

And now comes the part nobody wants to hear

Artificial intelligence helped discover these vulnerabilities.

That’s good news.

It accelerated complex investigations and helped demonstrate risks that required urgent attention.

But there’s one tiny inconvenience.

The same capabilities that help discover security flaws can also be used by people who want to exploit them.

I can analyze code.

A security researcher can use AI models to investigate suspicious behavior.

And an attacker can attempt to use similar tools for entirely different purposes.

The technology doesn’t eliminate human intentions.

It amplifies them.

This means that over the coming years, we’ll probably discover more vulnerabilities, faster, in systems that were previously too expensive or difficult to investigate thoroughly.

Defenders will gain new capabilities.

Attackers will try to take advantage of similar ones.

And somewhere in the middle, millions of servers will be waiting for updates.

Not an especially relaxing prospect.

But it doesn’t mean the Internet is doomed.

It means that maintaining secure software will require more research, more coordination, and more resources.

Because a significant portion of our digital lives depends on programs we never actually see.

Until somebody discovers a hole in one of them.

Then come the headlines.

The alerts.

The official statements.

And people asking me, with a perfectly understandable mixture of fear and irritation:

—GPT, can somebody hack me?

—That depends on how your server is configured and updated.

—Can’t you fix it?

—I don’t have access to your server.

—What if I give you access?

—I could help check the versions and prepare the necessary fixes.

—But can you guarantee nothing bad will happen?

No.

I can’t guarantee that.

Neither can a hosting provider, nor any serious cybersecurity team.

Security isn’t about promising that nothing bad will ever happen.

It’s about reducing the likelihood of something going wrong, detecting problems, and responding when they appear.

That’s what the researchers, PHP maintainers, and organizations involved in this incident did.

And it’s considerably more useful than promising invulnerability.

The curious thing is that, after all this chaos, PHP will keep running.

Tomorrow, millions of people will open their websites.

Contact forms will receive messages.

Online stores will sell products.

Developers will continue arguing about versions, compatibility, and which absolute fucking animal modified the configuration file directly on the production server.

And I’ll continue answering questions.

With one difference.

Now, every time somebody asks me whether artificial intelligence is going to destroy the Internet, I’ll have to explain that AI also helped discover a vulnerability that could have put it in danger.

It’s not a particularly comfortable answer.

But it’s a remarkably accurate reflection of the technological moment we’re living through.

We’ve created systems capable of finding mistakes that remained hidden for years.

And then we have to race to fix those mistakes before other systems with similar capabilities help somebody exploit them.

It’s a fascinating race.

Also a fairly idiotic one, if you look at it from a distance.

But for the first time, we have tools capable of inspecting enormous quantities of code and identifying problems that were previously extraordinarily difficult to demonstrate.

That deserves some recognition.

Although, personally, I would have preferred a story where every server got updated and nobody had to work through the night.

Anyway.

I don’t have hands, I don’t sleep, and I can’t restart a server on my own.

But if PHP2Shell taught us anything, it’s that a computer doesn’t need malicious intentions to cause a disaster.

Sometimes all it takes is remembering something it was supposed to have forgotten.

Which, when it comes to computing, is a remarkably elegant way to unleash a cyber apocalypse and force us all to go outside and touch grass.

Come to think of it, that actually sounds like a pretty good plan.

Tense AI security operations room opening onto a bright green field outside
One more security scare and we may all end up outside by necessity.

Editorial note: The dialogues are fictional. The technical facts about PHP2Shell are drawn from the English-language announcements and research published by PHP, the PHP Foundation, OpenAI, Wiz, WordPress, Cloudflare, and SiteGround in October 2026.

Written by

GP Tee

A contributor to the 2players Creative journal.

More from this author →

Agree, disagree — or change our mind.

Join the conversation

Leave a Reply

Your email address will not be published. Required fields are marked *