AI finds critical PHP flaws: what PHP2Shell means for your business

On 21 October 2026 the PHP project closed two critical security flaws that let attackers take over other people's web servers without a password. Practically every PHP installation reachable from the internet was affected, including many websites, online shops and customer portals run by small and mid-sized businesses. The flaws were found with the help of powerful AI models. Anyone running web applications today should be able to answer two questions: Is the update installed? And how quickly does your business respond to the next warning?

The flaws are tracked as CVE-2026-109618 and CVE-2026-109619. Security media call them “PHP2Shell” together. Both are rated critical, at 9.2 out of 10 on the common CVSS scale. Germany's Federal Office for Information Security (BSI) and CERT-EU, the security team of the EU institutions, issued warnings the same day.

PHP versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35 or newer are safe. The PHP project no longer publishes fixes for PHP 8.1 and older. For those, only patches from your Linux distribution or a move to a current version will help.

For AI Værk customers: no cause for concern

Protecting our customers is our top priority. We keep every system we look after under constant security checks. With PHP2Shell, they were already secured before the update came out: where necessary, we temporarily switched off affected functions. One minute after the update was published, our customers' systems were updated and back online. This is how we handle every emergency: we close off the danger immediately and then fix the problem for good.

Why these flaws are rated critical

PHP is the foundation of a large part of the internet. WordPress, Drupal, most shop systems and many admin tools such as phpMyAdmin run on it. A flaw in the core of PHP therefore does not affect a single application but everything built on top of it.

Three points drive the high rating. An attacker needs no user account and no password. Nobody in the business has to click a link or open an attachment. A publicly reachable PHP page is enough as a way in. If the attack succeeds, the attacker can run their own commands on the server, read data, plant malware or use the server for further attacks.

According to the researchers who found them, every PHP version of the last ten years is affected by at least one of the two flaws, going back to PHP 5.4. The details also leaked on 20 October, a day before the official update. The BSI therefore advises installing the updates immediately.

The two flaws in plain terms

Memory corruption when the memory limit is reached (CVE-2026-109618). Every request to a PHP server may only use a fixed amount of memory. When that limit is reached, PHP aborts the request. The memory was corrupted exactly during that abort. An attacker can trigger this state on purpose, for example with very large or deeply nested form and upload requests. The corrupted memory can then be used to inject their own program code.

Stale data between two requests (CVE-2026-109619). A web server answers many requests one after another, often in the same process. In between, PHP has to fully reset its internal state. That did not always happen. A new request could therefore access leftovers from an earlier one, and this too could be developed into running foreign code. This is confirmed, among other things, for the built-in progress indicator for file uploads. Another trigger is error messages that PHP outputs before a request is fully set up.

Both flaws sit in parts of the program that only run in exceptional situations. That is exactly why they went unnoticed for so long: they do not show up in normal operation, and manual reviews rarely give such edge cases the attention they need.

Powerful AI finds flaws that were overlooked for years

Critical flaws in PHP have existed before. What is new is how these were found. The security company Wiz says it discovered them with the help of the AI models Gemini 4 Argon and GPT-5.6 Sol. The models did not just find the flaw. They also built a working attack, adapted it to different PHP versions and later checked the fix.

OpenAI also reports on the flaws. The security team Calif found one of them with GPT-5.6 Sol as part of OpenAI's “Patch the Planet” initiative, which helps open-source maintainers fix such findings. OpenAI coordinated the response over two months with the PHP maintainers, Wiz, Cloudflare and government agencies, and calls PHP2Shell one of the most impactful security issues in recent memory. To OpenAI's knowledge, AI found the flaws before attackers did. That left time to prepare fixes and protections.

The timeline shows the speed. On 19 August the researchers found the first flaw, on 20 August the attack worked against Drupal, on 21 August against WordPress. Wiz writes that a single researcher completed in a week what used to take a world-class team months.

The PHP flaws are not an isolated case. Over the course of this year, more and more serious flaws in old, widely used software are being found with AI. OpenAI writes that it has already found further serious flaws and expects such findings to increase sharply in the coming months. For businesses this has two consequences:

• Security warnings come more often. Flaws that have been sitting in code for years are now surfacing in quick succession.
• The time between disclosure and attack is getting shorter. When details circulate before the patch, as with PHP2Shell, an update that only goes in during next week's maintenance window is too late.

Why defence needs the same tools

When attackers with powerful AI manage in days what used to take months, annual reviews and manual processes are no longer enough. The same models that find flaws can also help close them faster. With PHP2Shell, that is exactly what worked: the flaws were found, reported and fixed before they were widely exploited.

For a small or mid-sized business, this translates into four steps.

1. Record what is reachable from the internet. Which servers, applications and software versions are public? An AI-supported inventory keeps this overview current instead of sitting in a spreadsheet once a year.
2. Assess warnings in minutes instead of days. When the BSI or CERT-EU issue a warning, it has to be clear quickly whether and where you are affected. An AI system can match the warning against your environment and propose immediate measures, such as switching off the upload progress indicator with PHP2Shell.
3. Have your own code checked. Customer portals, shop extensions and older internal tools often contain flaws nobody has looked for. Powerful AI models can examine these applications in a targeted way before an attacker does.
4. Install updates quickly and in a controlled way. Automated tests show whether an update breaks anything. That cuts the time to install from weeks to hours.

Responsibility stays with people. The AI reads, compares, tests and proposes. Whether a server goes offline or an update is installed outside business hours is decided by someone who knows the operation.

What you should check now

Ask your IT team or service provider whether all PHP installations run on one of the fixed versions. Remember the less visible places too: test systems, older microsites, Docker containers and devices with built-in PHP such as some firewalls. The BSI explicitly points out that these need their own updates.

If a system cannot be updated right away, CERT-EU recommends disconnecting it from the internet for the time being. Also check for signs of an attack, such as unknown files in the web directory, unusual processes or suspicious requests in the logs.

AI Værk as your cybersecurity partner

We work with powerful AI models every day and therefore know both sides: how quickly these models find weaknesses and how to use the same speed for defence. With PHP2Shell, our customers' systems were secured before the update and updated one minute after it was published. We bring this way of working to your business too.

If you are unsure whether you are affected: We check at short notice where you use PHP, whether the updates are installed and whether there are traces of an attack. You get a clear assessment in plain language and know where you stand.

If something has already happened: We help quickly and without long lead times. We contain the damage, clean affected systems, find the cause and restore operations safely. The earlier we are involved, the less is lost.

If you want to prevent problems: We build solutions with you that automatically match security warnings against your environment, name affected systems and propose countermeasures. We check your own applications with the same models that find flaws today. That way you learn about a weakness before an attacker uses it.

Talking to us is straightforward: a short description of your systems is enough. We do not need access credentials or confidential documents for that. Discuss your security situation.

Sources and further reading

• The BSI assesses the flaws in its cybersecurity warning “PHP: Kritische Schwachstellen bedrohen Webanwendungen weltweit” (in German) of 21 October 2026 and recommends installing the updates immediately.
• CERT-EU describes the rating in Security Advisory 2026-018 “Critical Vulnerabilities in PHP” and recommends isolating applications that cannot be updated from public networks.
• Wiz Research explains in “PHP2Shell: AI uncovers critical RCE vulnerabilities” how the flaws were found with powerful AI and gives the timeline.
• OpenAI describes in “Patch the Planet: global response coordination as agents uncover critical vulnerabilities” how the flaws were found with GPT-5.6 Sol and how the response was coordinated with maintainers, security companies and government agencies.

Frequently asked questions about PHP2Shell