If anything has to follow Apple’s “it just works” philosophy to the letter, it’s the Internet. In fact, we don’t ask ourselves what’s wrong until something goes wrong. And yesterday things got very, very ugly. Two critical vulnerabilities in PHP, the language in which a good portion of the planet’s websites run, forced various European public services to disconnect from the network.
In this case it was not a new chapter of La Liga against Cloudflare, but something more serious in the framework of cyber security and it demonstrates how AI, in this case, can be a great ally in finding serious flaws in software before anyone else.
In short. The alarm went off at noon this past October 21. The PHP community released several versions that fixed two vulnerabilities (CVE-2026-109618 and CVE-2026-109619) that allowed an attacker to execute code on a server without logging in. When a non-malicious actor discovers such a vulnerability, it quickly recommends combating an organization directly to act as a cybersecurity issue. possible and consider taking the system offline.
The reason? Well, since the vulnerabilities affect how PHP handles certain memory errors during website service requests, there was a possibility that the server could be in an insecure state and hence, an attacker could take advantage of the code execution behavior remotely. That attacker has direct access to the keys to the servers that power more than half of the Internet
This translates to countries like Spain, France, Denmark, Finland or the Netherlands ‘throwing’ public pages and services (such as state security, education or administration pages) into a clear gesture of cutting corners. It is estimated that PHP is present on 69.3% of websites and perhaps the most recognizable example is WordPress which powers around 40% of all websites. It’s an insane number.
Tracker. That’s where artificial intelligence comes into play. Regardless of whether you like AI more or less, generative general intelligence must be distinguished in ways that can be used in research or cybersecurity. For example, OpenAI has a program called Patch the Planet that helps researchers find vulnerabilities in networks.
Security firm Wiz gave the first notice. Using both GPT-5.6 Sol and Gemini 4 Argon, he demonstrated remote code execution in WordPress, Drupal and Joomla. Within the Patch the Planet program, another firm called Calif expanded the list of vulnerable sites to Magneto, MediaWiki, NextCloud or phpMyAdmin. Of these seven, virtually the entire Internet relies on PHP.
Human role. Now, AI didn’t do the job alone. Researcher Ryan Austin from Calif explained that, to find out, he started from a known problem: a memory error that was reported to PHP in 2020 and detected again by an external audit in 2024. It is curious that this happens, but even if the flaw exists, it has not been possible to demonstrate that it can be exploited.
So Austin intentionally used a model to find ways to create those errors and, in the process, reviewed the code for vulnerabilities. He then asked the model to convert the pattern into a reusable instruction, which led to the second vulnerability. Califf said they didn’t use special tools or models that weren’t available to anyone, and that no model would have found all of these flaws unless people knew the history or intent of the attack.
For their part, the researchers at Wiz chained the failures and adapted the attack with models as support, ensuring that what a single researcher could do in a week required months and a top-level team. Of course, this makes it clear that adapting and configuring the attack to each version of PHP was a task for experts, although the models allow everything to be accelerated.
Cybersecurity as a goal for AI. Because both OpenAI and Anthropic have been running on cyber security for months in their sights. They are companies that want to make money and, furthermore, go public. When it seems that few people are willing to pay for such astronomically expensive technology, they must find other sources of income, and cyber security is the ideal scenario. Both continue to make their models more capable of finding critical flaws in systems and, if OpenAI with GPT-5 already points the way by the end of 2025, it’s Anthropic’s mythos that set off the alarm.
In its presentation, Anthropic commented that it has identified thousands of critical vulnerabilities in all operating systems and all browsers, including those that have been with us for decades in systems like OpenBSD or FFMPG. They put the model at the service of companies under Project Glasswing, and companies such as Mozilla commented that it was very helpful in patching critical bugs that they had not seen.
OpenAI went down the exact same path with GPT-6 Astra, noting that it achieved a 100% level in tests like Exploitbench compared to 78.5% for GPT-5.6 Sol, and that it was able to execute arbitrary code in browsers with all protections turned on. Mythos is not available in a commercial version and is offered to companies within strict yu security programs, although the GPT-6 Astra, the user version refuses to launch advanced offensive tasks, such as creating proof of concept exploits.
Old watch, new problem. And the underlying problem is that, even though it exploded a few hours ago, it was a ticking bomb. Days before the release of PHP patches and bugs, Amazon Threat Intelligence detected an attempt to exploit the PHP2Shell vulnerability on a network of honeypot servers. Amazon believes a malicious actor did this because information about the vulnerability has not yet been released, and they admit they don’t know how they got the information.
Cloudflare also reported an attack on its systems in an action that seemed related to what they saw from Amazon, and as I say, the problem is that it comes from a distance. According to Sansec (a firm of e-commerce security experts), since 2012 there has been vulnerable code related to PHP 5.4, from which PHP2Shell is derived. And we already said that the memory flaw known since 2020 and seen again in 2024 was ready for someone to exploit.
And if the question is how it is possible that it was never fixed, the explanation is as humane as possible: the PHP Foundation was fully aware of it, but the patching process was held up by a lack of consensus and resources.
No software is safe from finding vulnerabilities. Today it’s time to update PHP. Another day in the world of cyber security – INCIBE-CERT
And in Spain? If such a large number of web pages are vulnerable, it is assumed that in Spain, with .es domains, we have few. INCIBE-CERT tells us that they have been monitoring this vulnerability for several days and published a notification with security recommendations on the morning of October 21. They noted that, at least at the time, they had no information that an attack exploiting this vulnerability had occurred in Spain, and that, while it is certainly the worst breach in PHP’s history, it is still a very secure system.
Regarding the number of affected domains, they comment that “Spain will not be more or less affected than other countries”, but of the almost two million active .es domains, they estimate that about one and a half million use PHP and the vast majority are hosted on shared hosting where it is a company that will update for all clients, they do not have to do anything. Seeing the CCN-CERT page with the closure notice yesterday was not very encouraging.
what happens now. As they say from INCIBE, “Another day in the world of cyber security.” And yes, but the problem is, if cybersecurity firms and organizations have very powerful tools to be able to exploit vulnerabilities before the “bad guys” and patch them before putting a Band-Aid on the wound, malicious actors also have new systems they can employ to discover cracks.
We’ve already said that GPT-6 rejects the command to attack to find exploits, but the world of AI has not only closed source, but also open source tools and heavy models that are becoming more and more important and have a training base on which other people can continue to build. And, while cybersecurity has always been a cat-and-mouse race, with AI that race accelerates at a dizzying pace.
Picture | Mohammad Rahmani
In Xataka | Microsoft considers that “frontier models are not credible” and that AI agents must be considered an insider threat
