On Wednesday, governments started switching off their own websites — on purpose
Somewhere in The Hague, the team running the Dutch National Cyber Security Centre’s website did something that sounds like a joke with bad timing: they took their own cybersecurity website offline. The Port of Rotterdam’s site went dark. Overheid.nl, the Dutch government’s front door, closed for repairs. Across Europe, the closures stacked up — Finnish public web services, Danish municipal pages, French prefecture websites, and in the United Kingdom, nearly three dozen government services went temporarily unavailable, with the Office for Budget Responsibility and NHS Scotland among the names on the list — the first wave of precautionary shutdowns triggered by PHP2Shell.
No large-scale successful exploitation has been confirmed in public reporting — but the number of compromised installations remains an open question. What happened was something stranger: governments and operators in country after country made their own precautionary decisions to go dark before something worse happened. The trigger was PHP2Shell, two critical vulnerabilities in PHP, the programming language behind roughly 70 percent of websites whose server-side language is known, and the foundation underneath WordPress, Drupal, Joomla, and Magento.
What makes this story different from every previous scramble-to-patch is who found the flaws, and who is already trying to use them.
The flaw sat for over a decade. The exploit took a week.
PHP2Shell — tracked as CVE-2026-109618 and CVE-2026-109619, rated 9.2 out of 10 in severity — allows an attacker to execute code remotely on a server without logging in. The second flaw can expose process memory, the kind of memory that holds passwords. Affected versions reach all the way back to PHP 5.4, released in 2012. These vulnerabilities had been sitting in the floorboards of the internet for over a decade.
The PHP Foundation knew something was there. Its researchers had identified a theoretical flaw as far back as 2024 but could not work out how it might actually be exploited — the theory sat on a shelf, unweaponized. The record does not explain why that leap stalled; what it documents is that human-led, AI-assisted research then produced working exploits. Then the Foundation joined OpenAI’s “Patch the Planet” initiative in July 2026, and the shelf got very short. Researchers at the security firm Calif reported the first flaw on August 11, using OpenAI’s GPT-5.6 Sol model. Wiz, the cybersecurity firm owned by Alphabet, reported the second on August 24, with its researchers using Google’s Gemini 4 Argon, OpenAI’s GPT-5.6 Sol, and its own Wiz Atlas tooling.
Wiz put the acceleration in startling terms: “With frontier AI models and the right tools, a single researcher completed in a week what previously would have taken a world-class team months of work.”
A note of precision is owed here, because the headlines saying “AI found the flaw” deserve scrutiny. These were human-led investigations: researchers directed the models, framed the questions, and validated the results. That comparison is the reported fact; two caveats matter. Discovery acceleration is not new: static analysis, fuzzing, and symbolic execution have been compressing security research for decades, and frontier models may be the next step on that curve rather than a break from it. And the ecosystem, for its part, functioned visibly — vendors coordinated, the PHP project shipped patches, Cloudflare deployed mitigations, governments issued warnings and acted. The week-versus-months comparison is reported; whether it marks a change of kind is an open question, not a verdict.
The other edge of the same blade
The choreography of responsible disclosure held for roughly two months — reports in August, a fix in the works. Then, on October 20, it broke: basic details of the vulnerabilities leaked on social media, and Calif mistakenly published a video demonstrating how the flaws could be exploited. The video spread on hacking forums, and the PHP project brought the October 21 release forward.
And here is the part that should worry you more than the shutdowns: within hours of disclosure, attackers were already adapting. Cloudflare, which reported observing attack traffic against its customers and deployed its own mitigations, also observed attackers mutating their exploit payloads with AI assistance to bypass those protections, tracking the volume of blocked matches and cross-referencing it against the rule logic. The defenders’ instrument, in other hands, probing the same crack within hours.
This is the symmetry at the heart of PHP2Shell: measurable compression on the discovery side — Wiz’s reported one week against a months-long baseline — and, on the attack side, the same class of tools already in hand, with AI-assisted payload mutation observed within hours of disclosure but no comparable baseline. Whether the shutdowns reflected doubt about a narrowed defender margin is not established; each institution’s rationale remains an open question. What they establish is a documented willingness to spend availability against an unquantified risk.
The warning, and the darkness
The U.S. National Security Agency issued a rare public warning: patch immediately, or pull vulnerable systems offline. The advice was echoed by authorities from Japan to the European Union, and by the United Kingdom’s National Cyber Security Centre. A platform affiliated with China’s Ministry of Industry and Information Technology warned of “two critical remote code execution vulnerabilities with severe impact.” Türkiye’s Presidency of Cybersecurity launched nationwide assessments and circulated an emergency advisory through its Cyber Incident Response Teams platform to government bodies, private companies, and critical infrastructure operators.
Read that NSA guidance again, because its bluntness is the whole story: patch immediately, or go offline. No large-scale successful exploitation has been confirmed in public reporting. The darkness was precautionary — operators interrupted their own services at real cost rather than leave them exposed. The Dutch cyber agency taking down its own site is the emblem of the moment: the people whose job is internet security chose interruption over exposure, for a little while — the cost, documented; the calculus behind each decision, not.
The question PHP2Shell poses
But the question is now unavoidable, and it will outlive this patch cycle: what happens when AI speeds both sides of the race? The internet’s most critical infrastructure — the languages, libraries, and protocols everything runs on — has been audited with steadily improving tools: fuzzers, static analysis, and now frontier models. The established facts are the warnings, the precautionary shutdowns, and their availability cost; the strategic implication is the question, not the conclusion. And if models could surface what a 2024 theory couldn’t crack, the reasonable question is how many similar dormant flaws sit in equally old codebases — and who finds them first.
On Wednesday, institutions answered by going dark, spending their own availability against an unquantified risk. Whether that bought real protection is unproven — no confirmed large-scale exploitation is not proof it worked, and the count of compromised installations stays open. What is established is the cost, and the question it leaves behind: what does defense look like when discovery compresses and attackers hold the same tools?
