PHP2Shell: AI-Found Critical Flaw Sparks Emergency Patching as Governments Take Websites Offline

Governments across Europe and Asia took websites offline on Wednesday as security teams raced to patch two critical vulnerabilities in PHP, the technology that powers as much as 70 per cent of the world’s websites. The flaw, dubbed PHP2Shell, was uncovered with the help of artificial intelligence — and security officials warn that the same AI capabilities could now put an unprecedented share of the internet at risk.

The US National Security Agency issued an unusual public warning on Wednesday, urging every organisation running PHP to install the available patches immediately — or to take their websites offline if they could not update right away, according to reporting by Reuters and other outlets. The FBI issued a parallel warning that the flaws could allow unauthorised access to servers and data or enable remote code execution.

Several government and commercial websites went dark briefly on Wednesday as administrators installed fixes, according to CNN. The countries taking the most visible action included Germany, where the digital ministry said it had taken some websites offline as a precaution, and France, where the interior ministry disabled official websites serving local administrative offices. Authorities in Belgium, the Netherlands, Finland and the Czech Republic also said they had shut down websites to install the patch, while the European Union’s cybersecurity agency advised applying the security patches immediately — or taking websites offline if updates were not feasible, according to Reuters.

In the Netherlands, government portals, the Port of Rotterdam and major hospital websites went offline, while public networks and school platforms in Finland and Denmark were also disabled. In Belgium, government agencies, parliament, school platforms and major media websites ceased operations temporarily. In Japan, Prime Minister Sanae Takaichi issued a public warning about the PHP risk on social media, and several Japanese services — including childcare centre systems and online games — were taken offline as a precaution.

In Australia, the Australian Signals Directorate’s Australian Cyber Security Centre issued a critical alert, its highest rating, telling organisations to find every instance of PHP running in their environments. The agency acknowledged that its own website, cyber.gov.au, was among the Australian government sites affected.

What PHP2Shell is — and why it matters

The vulnerabilities were assigned the identifiers CVE-2026-109618 and CVE-2026-109619, and both carry a CVSS severity score of 9.2 out of 10, according to the technical reporting collated by CXO Digital Pulse. They represent a class of problems in the way PHP manages a computer’s memory — weaknesses that sit inside the PHP runtime itself. That distinction matters: updating a website’s content-management system, application code or plugins alone is not enough. The underlying PHP installation has to be replaced.

According to the Wall Street Journal, if exploited, the flaws could allow attackers to take control of any system running PHP software — stealing information or launching cyberattacks on anyone visiting a hacked website. The vulnerabilities can be exploited in many ways, the Journal reported, some requiring just a few seconds on smaller systems, though more complex servers may take hours to compromise.

The PHP security team released patched versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, and administrators were advised to install the appropriate update, restart their PHP workers, and examine any systems that had been exposed before patching. Organisations unable to apply the fixes immediately were told to isolate vulnerable applications or temporarily remove them from public networks. Patched source code was also released for older PHP versions 7.4, 8.0 and 8.1, and the Linux vendors Red Hat and AlmaLinux backported fixes to retired PHP streams, giving older systems a path to patch that had not initially existed.

The patching effort — and the leak

The road to disclosure was an exercise in secrecy. According to the Wall Street Journal, the flaws were first uncovered in August by Ryan Austin, a researcher at the cybersecurity company Calif, during an audit of PHP’s source code with the help of GPT-5.6 Sol, one of OpenAI’s most advanced models. OpenAI funded that audit and others, part of an effort by AI companies to help open-source projects address vulnerabilities now being surfaced by AI systems.

The class of bugs had been hiding in PHP’s source code for years. As the Journal reported, a researcher had flagged the problem area decades earlier, but the flaw had been considered too difficult to exploit and was largely forgotten — until AI made it practical. Austin figured out how to control the code running in a computer’s memory with as few as a few dozen requests to a server. On August 11, Calif’s chief executive Thai Duong called Volker Dusch, a security engineer with the nonprofit foundation that manages PHP. “There’s a serious issue here,” Duong said.

An informal working group formed among security researchers and software engineers, collaborating with software distributors and security companies to build the patches. Researchers from Google’s Wiz cybersecurity team reached out with exploits they had discovered independently of Calif, and OpenAI’s own research team contributed findings. The patching effort was massive: close to 200 files were fixed, and developers wrote more than 10,000 lines of code during the process.

The patch date was set for October 22. Then, as the Journal reported, the news leaked: on Tuesday, a post to Hacker News broke the secret just under two days before the planned release. The PHP team rushed the patch out and warned the world on Tuesday to stand by and expect a critical security update on Wednesday. Dusch described the approach as “a good balance to inform defenders, and not give attackers too much time to develop exploits.”

Wiz researchers described PHP2Shell as among the most severe vulnerabilities ever identified in PHP — and arguably in software overall. The flaws had existed for decades, Wiz wrote, but demonstrating that they could be exploited required a sequence of steps approaching the practical limits of human researchers. “Frontier AI models substantially reduced that barrier,” the researchers wrote, “turning a longstanding flaw into an exploit.” Wiz’s co-founder Ami Luttwak called it “the most severe AI vulnerability that was ever found,” and one that was likely too complex to have been discovered without AI.

Attacks are already being attempted

With the technical details now public, the race is on. Web infrastructure provider Cloudflare, which worked with the PHP Foundation, Calif, OpenAI and Wiz on the response, said it began deploying web-application firewall protections for its customers before the public disclosure — and subsequently observed attack traffic targeting the vulnerabilities.

Security vendors have begun reporting real exploitation. Hadrian, a security vendor, said it was seeing PHP2Shell exploited at scale, mostly through PHP’s session upload progress path. Patchstack said it had confirmed multiple active exploitation incidents. At the same time, the picture is mixed: security firm Huntress said it had not observed exploitation across its partner systems as of late on October 21, and the US cybersecurity agency CISA said the same in its alert that day. Neither flaw had been added to CISA’s Known Exploited Vulnerabilities catalogue as of late on October 21, and no victim organisation had been publicly named.

Echoes of Heartbleed

The comparisons have been immediate and deliberate. Dusch, the PHP security engineer, told the Journal that PHP2Shell is the most severe security issue PHP has faced in more than a decade, likening it to Heartbleed — the 2014 vulnerability in encryption technology estimated to be used by two-thirds of the web, which forced a global race to protect hundreds of thousands of servers. Wiz researchers made the same comparison: Heartbleed, CNN noted, allowed cybercriminals to access data that should have been protected, and like Heartbleed, PHP2Shell could serve as a turning point in cybersecurity.

Analysis: Why It Matters

1. The first AI-found crisis at internet scale

PHP2Shell is not just a bad bug — it is a historical marker. This is the first vulnerability found with AI’s direct assistance to threaten a majority of the world’s websites simultaneously. Wiz reported that a single researcher completed in a week what previously would have taken a world-class team months, using models including Google’s Gemini 4 Argon and OpenAI’s GPT-5.6 Sol. That acceleration is the real story: the same class of models is now available — legally or otherwise — to the attackers on the other side. When the discovery tools and the exploitation tools are the same tools, the defender’s head start shrinks from months to hours. The coordinated disclosure system that held this secret since August will not always be lucky enough to find the flaw first.

2. Seventy per cent is not a statistic — it is everything

When researchers say PHP powers up to 70 per cent of websites, they are describing the invisible plumbing of modern life: government portals where citizens file taxes and book medical appointments, hospital sites, school platforms, small-business storefronts, charity donation pages, the websites of local newspapers. The take-down lists published on Wednesday — port authorities, hospitals, parliaments, childcare systems — show how deep that plumbing runs. Unlike Heartbleed, which struck the encryption layer, PHP2Shell strikes the application layer, where the crown jewels (databases, logins, customer records) actually live. That is why the NSA’s warning was so blunt: this is not an abstract risk to be managed at leisure; it is a take-the-site-offline risk.

3. The defenders have a new, terrible arithmetic

The traditional security model assumes defenders get months to patch while attackers take months to weaponise. PHP2Shell breaks both ends of that bargain. The patch required close to 200 file fixes and over 10,000 lines of code — an enormous, delicate update that system administrators must test before deploying, because a bad PHP upgrade can take a site down as surely as an attack. Meanwhile, Thai Duong’s warning hangs over the whole episode: “Some dude on Twitter could take the patches and run AI against it and post some exploit script. All hell will break loose.” The patch itself is a treasure map. Every hour a site stays unpatched after Wednesday is an hour someone else is feeding the patch into an AI and asking for the exploit.

4. Shutting down the internet’s waiting room

There is a quiet irony in the most dramatic detail of this story: governments chose to switch off rather than risk it. The ASD’s own cyber.gov.au — the very site telling Australians to patch — was itself affected. France’s interior ministry disabled the portals citizens use for local administration. Belgium shut down its parliament’s site. These are not organisations that chose downtime lightly; they chose it because the alternative was to leave the front door to citizen data standing open. That choice reveals an uncomfortable truth about critical infrastructure built on shared open-source software: there is no private hotline, no premium support tier that matters when the flaw is in the foundation. The only lever the state had was the off switch.

5. The policy lag is now the biggest vulnerability

Notice what happened on Wednesday: the NSA posted a warning on social media. Japan’s prime minister posted on social media. The EU’s cybersecurity agency advised website operators through its usual channels. This is emergency infrastructure coordination being conducted through tweets and blog posts. The tools for a coordinated global patching response at this speed barely exist — CISA’s Known Exploited Vulnerabilities catalogue, the industry’s main must-patch-now list, had not even listed the flaws as of late on October 21. The attackers, meanwhile, need no coordination at all: Hadrian and Patchstack were already seeing exploitation attempts within hours of disclosure. If PHP2Shell is the template for the AI-discovered vulnerability era, the institutions built for the human-paced era are already behind.

What to Watch Next

1. Exploitation reports in the next 48–72 hours. The split between vendors seeing active exploitation and agencies not yet confirming it will resolve quickly — watch for the first named victims and whether CISA adds the CVEs to its Known Exploited Vulnerabilities catalogue.
2. Patch uptake across shared hosting. The hardest-to-reach systems are the millions of small sites on shared hosting and the older PHP streams that only got backported fixes — the long tail is where attackers will hunt.
3. The Hacker News leak investigation. How the patch plan leaked on Tuesday, and whether it gave attackers a head start before Wednesday’s disclosure, will shape how future coordinated disclosures are managed.
4. Regulatory follow-through. Whether the EU, the US and other governments move from emergency advice to mandatory patching timelines or new open-source security funding — and what the Australian, Singaporean and Japanese agencies report next.
5. The AI angle deepens. Full technical write-ups from Wiz, Calif and OpenAI on exactly how the models found the flaw will set the template for both sides — and for the debate over AI safety that has been building from Washington to Silicon Valley.

Sources

• The Wall Street Journal — “AI Just Found a Bug That Puts up to 70% of the World’s Websites at Risk”
• CNN — “Global websites shut down after AI found a critical flaw affecting 70% of the internet”
• CXO Digital Pulse — “PHP2Shell Vulnerabilities Trigger Emergency Patching Across Global Web Infrastructure” (CVE-2026-109618 / CVE-2026-109619, patch versions)
• Insurance Business — “Critical global website flaw sparks alerts across Asia-Pac as attacks begin” (ASD, CISA, exploitation reports)
• KuCoin News — “Global Authorities Warn of Critical PHP Vulnerabilities, Take Websites Offline” (NSA/FBI warnings, European shutdowns)