AI Found the PHP2Shell Flaw Before Hackers Did. Someone Still Got There Four Days Early

OpenAI, Wiz and Calif used AI to uncover two critical PHP vulnerabilities that knocked government websites offline from Japan to the Netherlands. Amazon says an exploit attempt came four days before the fix. Here is what that means for AI security.

On Wednesday night in Tokyo, Japan's prime minister posted a request on X that you rarely hear from a head of government. "We have an urgent request for everyone to take action tonight," Sanae Takaichi wrote.

The emergency wasn't a storm or an earthquake. It was a programming language.

PHP quietly runs about 70% of the world's websites, according to the Associated Press. On 21 October, its maintainers shipped fixes for two critical flaws, CVE-2026–109618 and CVE-2026–109619, now known as PHP2Shell. A single request from a stranger, with no login, could let them run their own code on a server before the website's code even starts.

Within hours, governments started pulling the plug on their own websites. On purpose.

The day the lights went off by choice

The Dutch national cyber security centre took its own site offline as a precaution, and others followed, including the Port of Rotterdam and the government portal Overheid.nl, Reuters reported. Finland, France and Australia temporarily took some government services offline to patch, according to The National. Wikipedia's operator took some systems down. In Japan, MUFG said some webpages and services were temporarily suspended. Bangladesh Bank, the country's central bank, took its website offline on 22 October after a global security alert, its spokesperson confirmed to The Daily Star.

Japan's chief cabinet secretary, Minoru Kihara, summed up the mood in three words: "Just to be safe."

The US FBI went further than most. It recommended that operators take sites offline if they couldn't update straight away. CERT-EU told organisations that couldn't patch to consider isolating internet-facing applications.

Costin Raiu, co-founder of TLP:Black, warned of the risk of "mass compromise of every vulnerable web server on the internet." Wiz, one of the teams behind the discovery, said the findings "rank among the most severe vulnerabilities discovered in PHP, and arguably in software history."

Who found it: the machines

This is the part that makes PHP2Shell different from the big bugs of the past.

The flaws were found by teams at OpenAI, Google-owned Wiz and a security firm called Calif, according to ABC7 and the Associated Press. Wiz researchers used AI models from Google and OpenAI during the work. The bugs were discovered in late August and kept quiet while a fix was built.

OpenAI called it "one of the most impactful vulnerabilities in recent memory" and said it "marks a new phase, where AI can find security bugs with internet-scale impact." It said frontier AI found the flaws "ahead of attackers, giving defenders time to prepare."

That is the good news. AI read through a codebase that has been around for decades and spotted what humans had missed.

Who else found it: we don't know

Now the less comfortable part.

On 20 October, a day before the planned patch, a user on the Hacker News forum posted the two CVE numbers, which were meant to stay secret until fixes were out, Infosecurity Magazine reported. The post was deleted. PHP's security team had already warned that it reserved the right to release early if details went public.

Then Amazon dropped a bigger surprise. Its threat intelligence team said its decoy servers recorded a PHP2Shell exploitation attempt on 17 October, four days before public disclosure, from an address tied to a commercial anonymisation service. Amazon said the timing suggests a threat actor, and that how that actor got the details remains unknown. Amazon was also careful to say it is "unclear if any of these attempts would have succeeded in real-world systems."

After the announcement, the floodgates opened. Cloudflare's chief executive, Matthew Prince, reported "millions" of exploitation attempts. Amazon's chief information security officer, CJ Moses, said: "PHP2Shell is now being broadly exploited in the wild."

Yet as of 22 October, officials in Germany, the Netherlands and Japan had not confirmed a successful break-in using the flaws, The Next Web reported. Germany's BSI said it had seen attacks but no successful use. Thomas Caspers, its vice-president, said: "It seems a little bit more complicated than expected to exploit."

Lots of knocking. No confirmed burglary yet. And one early visitor whose identity and source nobody can explain.

The window is shrinking

AI is now good enough to find a flaw like this. That means it can also help whoever else is looking. In this case a leak, a mystery early attempt and published proof-of-concept code all landed inside the same week.

Lukasz Olejnik of King's College London put it plainly: "The scale is huge, and with AI, exploitation may happen instantly."

There is also a supply chain sting in the tail. Palo Alto Networks said all PAN-OS and Panorama versions are affected, and fixed releases had not shipped when Help Net Security reported it. Many organisations will be running PHP inside products they never think of as PHP.

Questions for the boardroom this week

Do we know every place PHP runs in our business? Including inside vendor appliances, plug-ins, marketing sites and that old microsite nobody owns.

How fast did we patch, and can we prove it? A dated record of what was patched and when beats a reassuring email.

Who decided whether to go offline? Governments made that call in hours. Make sure someone in your business has the authority to do the same.

Are our suppliers telling us their exposure? Ask for it in writing, product by product.

Are we using AI to find our own flaws first? If attackers can point AI at your code, you should be doing it before them, with someone independent checking the results.

Evidence, not assurances

PHP2Shell is a rare story where AI looks like the hero and still leaves a mess. The same capability that found the bug early shortened everyone's reaction time. For every organisation running PHP, the hard question on Wednesday was simple. Could they show, quickly, where they were exposed and what they had fixed?

At Web Dance, that idea sits behind AISOF, the AI Integrity and Secure Operations Frameworks we own and issue. Outcomes are decided by an approved assessor, never by the organisation itself.

AI found the hole first this time. Next time, the first one there might not be on our side.

---

Web Dance Pty. Ltd is an Australian company based in Sydney. We own and issue AISOF®, the AI Integrity and Secure Operations Frameworks, a family of assurance frameworks covering organisational AI governance, named AI products and five sector specific tracks. AISOF is a registered trademark of Web Dance Pty. Ltd, registered with IP Australia. More at aisof.org.