AI Just Found Critical Internet Flaws. What Comes Next Is the Real Story
AI models uncovered two severe PHP vulnerabilities affecting millions of sites. Here's what the "defenders window" means for future cyberattacks.
What happened with PHP, in plain terms
In October, researchers using AI models found two critical flaws inside PHP, the programming language that runs an estimated 70% of websites on the internet, including most WordPress sites. The PHP team rated both bugs 9.2 out of 10 for severity, the kind of score that lets an attacker take over a server remotely, without a password or an account. Security company Wiz found an early version of the problem using AI back in August. A security firm called Kaliff found one of the flaws using OpenAI’s GPT-5.6 Soul model. OpenAI’s own team found a more complete version in September. Three separate groups, using AI tooling, converged on the same hidden weakness in software that has existed for roughly 30 years.
What followed was a two-month quiet period where PHP maintainers and security researchers worked on a fix before telling the public, standard practice for responsible disclosure. The plan fell apart when details leaked early. PHP pushed the patch out on October 21st, ahead of schedule, and the response from governments and agencies was immediate: the NSA and UK’s National Cyber Security Centre told organizations to patch immediately or take systems offline, the FBI issued a “patch and hunt” alert, and the Dutch, Finnish, French, and Australian governments reportedly pulled some of their own services offline. Japan’s prime minister posted directly on X urging citizens to check their websites that same night.
TL;DR
- Two critical PHP flaws, rated 9.2/10 for severity, were found by multiple teams using AI models including GPT-5.6 Soul and Gemini, not by a single human researcher working alone.
- The fix leaked early, forcing PHP to release its patch ahead of schedule and triggering emergency alerts from the NSA, FBI, and UK’s NCSC within hours.
- A large share of the internet runs unmaintained software: one analysis of 7.2 million PHP sites found roughly 46% running end-of-life versions that may never get this patch.
- OpenAI calls this discovery part of a “defenders window”, the idea that AI can find dangerous bugs before attackers do, but that window only holds if the details stay contained, which didn’t fully happen here.
- This is one of several vulnerabilities already found, with OpenAI stating it expects the pace of AI-discovered flaws to increase sharply in coming months.
- A separate incident, the eMudhra certificate compromise, showed a trusted digital certificate authority using a private key as simple as the number 4, underscoring how uneven security practices are even in supposedly trusted infrastructure.
- Smaller, unmaintained websites and organizations outside major security partnerships are the most exposed, since they’re slowest to patch and least likely to hear about the problem at all.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How does AI find vulnerabilities faster than humans?
Finding a security flaw is closest to proofreading a book that’s millions of pages long, looking for one sentence that’s subtly wrong only under specific conditions. Humans are capable of this work but get tired, lose focus, and can only review so much code in a day. AI models don’t have that ceiling. They can run the same check against the same codebase repeatedly, test it against many hypothetical conditions, and run many parallel copies of the same analysis at once.
In the PHP case, researchers at Kaliff documented targeting “interrupted work,” specifically memory cleanup processes that run when PHP runs out of memory mid-operation. That’s a narrow, technical corner of the software that would take a human team a long time to notice. Once researchers found one instance of the pattern, they reportedly had the AI generalize the technique into a reusable method for finding similar bugs elsewhere, a self-improving approach where each discovery sharpens the tool used to find the next one. That compounding effect, not just raw speed, is what makes AI-assisted vulnerability hunting fundamentally different from traditional manual auditing.
What is the “defenders window” and why does it matter?
OpenAI used this term to describe the period of time between when a flaw is discovered and when attackers find out about it. If AI helps defenders find a bug first, that window gives maintainers time to build and ship a fix before anyone can exploit it. OpenAI said that, to its knowledge, its model found these PHP flaws ahead of attackers, calling it one of the most impactful vulnerabilities in recent memory and describing it as marking a new phase where AI can find bugs with internet-scale impact.
The problem is the window is fragile. In this case, researchers and maintainers spent roughly two months quietly preparing a coordinated fix, which is exactly how responsible disclosure is supposed to work. Then details leaked before the patch shipped, forcing an early release and giving attackers a head start they weren’t supposed to get. Security firm Cloudflare said it was already seeing early attacks within hours of the leak. Even in a best-case scenario, with multiple serious organizations cooperating on a careful timeline, the defenders window still slammed shut early. That’s the uncomfortable lesson: finding the bug first isn’t enough if the process around disclosure isn’t airtight.
Why is this only the beginning?
Two things make this moment different from a routine bug disclosure. First, OpenAI stated this was one of several significant vulnerabilities its systems have already found and are working to fix, and that it expects similar findings to increase sharply in the coming months. This wasn’t a one-off; it’s a preview of a pipeline.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Second, the model that found these flaws, GPT-5.6 Soul, is already a generation behind OpenAI’s current frontier models. Commentary circulating after the disclosure pointed out that the industry had already moved on to newer models by the time this vulnerability became public. If an older model can surface a critical flaw in infrastructure that millions of sites depend on, the open question is what newer, more capable models will find, and how much faster they’ll find it. The same capability that lets AI patch the internet faster also lowers the skill and time required to discover exploitable flaws in the first place, which is why this story cuts both ways: it’s good news for defenders only as long as defenders keep winning the race.
Who is most at risk from AI-discovered vulnerabilities?
Not everyone is equally exposed. Large, well-resourced organizations with dedicated security teams, and companies already participating in coordinated AI-security initiatives, tend to patch fast and have the staff to monitor for exploitation. The risk concentrates heavily in smaller, less-resourced systems: old online stores, community forums, local council or nonprofit websites, and any service set up once and left unmaintained.
An analysis cited during the PHP disclosure found that of the websites whose PHP version could be identified, close to half were running versions that no longer receive official security updates at all. Those sites simply won’t get this fix, regardless of how good the patch is. The people running them are often not reading cybersecurity alerts and may not learn about the vulnerability for months, if ever. That’s where the real consumer risk sits: not in large platforms, but in smaller sites that hold personal data (names, emails, old passwords, sometimes payment details) and have nobody actively watching for a breach.
A separate, related incident involved eMudhra, a certificate authority trusted by browsers to verify that websites are legitimate. Reports indicated some of its cryptographic keys were far weaker than the extremely long random numbers normally required, undermining a system millions of browsers rely on to confirm a site is who it claims to be. It’s a reminder that trust infrastructure across the internet varies wildly in quality, and weak links can sit in unexpected places.
What can people actually do about it?
Basic hygiene matters more now, not less. Use unique passwords for every account so one breach doesn’t cascade into others. Turn on two-factor authentication wherever it’s offered. Keep browsers, operating systems, and apps updated, and run reputable antivirus and anti-malware software on all devices, including phones. Be cautious with links in texts and emails during periods when a widely publicized vulnerability is actively being exploited, since scammers move fast to exploit public attention on a breaking story. If you run a website on PHP or any affected software, confirm the patch is actually applied, don’t just assume an update notification means it happened.
For data already exposed through old accounts on sites you no longer use, there’s no retroactive fix. Assume any password reused across multiple old accounts is compromised and rotate it.
Frequently Asked Questions
What is PHP and why does this vulnerability matter so much?
PHP is a programming language that powers the back end of a large share of the internet, estimated at around 70% of websites, including most WordPress sites, online shops, forums, and booking systems. A flaw in PHP itself, rather than in one individual website, affects every site built on top of it simultaneously.
Did AI cause this vulnerability or find it?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
AI models found the vulnerability; they didn’t create it. The flaw had existed in PHP’s code for an unknown period of time. Multiple research teams used AI models, including OpenAI’s GPT-5.6 Soul and Google’s Gemini, to analyze PHP’s codebase and surface the bug faster than manual auditing likely would have.
What is the “defenders window” concept?
It refers to the time between an AI system discovering a security flaw and that information reaching attackers. The goal is to use that window to build and deploy a fix before bad actors can exploit the bug. In the PHP case, the window was supposed to last through a coordinated two-month disclosure process but closed early when details leaked.
Why are so many websites still unprotected even after a patch exists?
Patches only protect software that’s still actively maintained. A large portion of PHP websites run end-of-life versions that no longer receive security updates at all, often because the site was built once and never updated again. Those sites remain exposed indefinitely unless someone manually intervenes.
Should regular internet users be worried about this?
Yes, in an indirect way. Even people who’ve never written a line of code can be affected if a site they signed up for years ago gets breached, exposing old passwords, emails, or personal details. The practical defense is using unique passwords per site, enabling two-factor authentication, and treating unexpected links or messages with extra caution during periods of active exploitation.
