Two critical flaws in PHP, the web language embedded in millions of sites, triggered a global defensive scramble on Oct. 21. The PHP project shipped emergency fixes after researchers said artificial-intelligence tools helped expose bugs that had survived for years. Government agencies, schools, hospitals and other operators temporarily took sites offline. The evidence supports a serious patching emergency; it does not show that every darkened site was breached or that an AI system independently hacked the web.
What happened on Oct. 21
The disclosure became a software-supply-chain event, not a single website incident. The PHP project released fixes for two critical vulnerabilities, and CERT-EU recorded the release at 13:30 UTC in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Both flaws were rated CVSS 9.2 and can allow unauthenticated remote code execution in affected deployments, according to the agency's security advisory.
PHP is often the unseen runtime behind public-facing services. Reuters reported that it supports large parts of the web and is built into WordPress, Drupal, Joomla and Magento. Its report described precautionary outages involving government services in the Netherlands, Finland and France, plus school, hospital and other public-sector sites. Operators were trying to create time to patch, not announcing that each site had been penetrated. Reuters' account records both the disruption and uncertainty about what attackers had achieved.
Once technical details and early proof-of-concept material circulated, CERT-EU said the likelihood of exploitation increased. It advised organizations to patch immediately or, where that was not feasible, isolate internet-facing applications until a fix was applied and verified. Those are risk-reduction decisions, not proof that a particular outage followed a successful intrusion.
What the two PHP flaws allow
Two CVEs, one core interpreter problem
CVE-2026-109618 concerns memory corruption during out-of-memory handling. An interrupted request can leave internal data partly initialized, after which cleanup may treat attacker-influenced memory as valid. CVE-2026-109619 concerns stale global state: information from one request can remain associated with freed memory and later be reused by another request. CERT-EU says both paths can lead to remote code execution without authentication. The PHP Foundation describes the bugs as failures in memory management and error handling beneath individual web applications.
The exposure is broader than a vulnerable plug-in. CERT-EU lists PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed among affected ways to serve web requests, with mod_php identified by the PHP team as the highest-risk setup. Versions before 8.5.12, 8.4.27, 8.3.36 and 8.2.35 are affected. End-of-life branches such as PHP 8.1 and older no longer receive upstream updates, although some vendors may provide backports.
A patch fixes the cause; isolation buys time
CERT-EU says only an update fully remediates the vulnerabilities. Network isolation can reduce exposure while a change is tested, but it can make services unavailable and does not repair the code. Affected products may also embed PHP without exposing it as a separately managed installation. Inventories therefore need to cover hosting platforms, distribution packages, containers and vendors that ship a PHP runtime.
The PHP Foundation said the release followed an embargo disruption after CVE identifiers circulated early. The project moved its release date while distributors prepared packages and acknowledged that limited specialist capacity and a slow security decision process had delayed work on a deeply rooted problem. Once the full impact was understood, maintainers and distributors worked toward a synchronized patch release.
What AI changed in vulnerability research
AI enters this story as an accelerator for human security research. Reuters reported that Calif and Wiz separately said AI helped them find the PHP flaws and that the firms coordinated with the PHP team and OpenAI. Wiz says its researchers used Wiz Atlas, Google's Gemini 4 Argon and OpenAI's GPT-5.6 Sol, crediting Gemini with a particularly significant contribution to finding the vulnerability, developing an exploit, detecting attacks and validating the patch. Those are a company's claims about its own work, not an independent measurement of model performance. Wiz's research post is primary evidence of what the team says it did, not proof that a model acted alone.
The technical reason AI mattered is practical. Memory-safety failures can require a researcher to trace allocator behavior, reproduce crashes, adapt assumptions across versions and test many changes. Wiz says models helped analyze failures and adjust an exploit across configurations. That may reduce the time and specialist effort needed for discovery and validation. It does not remove controlled testing, human judgment or responsible disclosure, and it does not establish that a demonstrated chain works against every PHP installation.
The PHP Foundation supplies a counterweight to an autonomy narrative. It says the project uses AI-assisted scanning and validation, but a person makes every call and works on every patch. The defensible description is therefore “AI-assisted vulnerability discovery and exploit research,” not “AI hacked websites.”
Why organizations went offline
Taking a site offline is a blunt response when a pre-authentication flaw sits in software that processes web requests before application logic can help. A public patch gives defenders a route to safety, but also gives attackers a target to study. Reuters reported that Amazon detected an attempted exploit against a honeypot on Oct. 17 and saw more attempts around disclosure; Amazon said it remained unclear whether they would have worked against real websites. Attempts and circulating proof-of-concept code describe a threat environment, not verified compromise of every affected server.
The trade-off was visible in the outages. Finland's state IT provider said public network services were taken offline as a precaution, while France's Interior Ministry put its sites into preventive maintenance, according to Reuters. Large providers can stage updates and inspect logs; small organizations on shared hosting may not control the underlying PHP process. Reuters also reported Automattic's statement that its hosted solutions and Tumblr were safe. That is provider-specific, not a general exemption for PHP-based sites.
What the response means for the web
This episode is a test of software maintenance as much as of AI. PHP's ubiquity lets a core interpreter flaw cross organizational boundaries: a site owner may manage content while a host, Linux distribution or platform vendor manages the runtime. CERT-EU's advice to update supported branches, consult vendors for backports and isolate systems that cannot yet be fixed reflects that chain of responsibility.
It is also a warning about disclosure timing. The PHP Foundation said the project moved its planned release forward after the CVE numbers leaked while distributors were still preparing packages. Premature disclosure can leave some defenders with a patch and others with only a vulnerability description. The Oct. 21 release balanced those risks, but did not guarantee that every ecosystem participant could update immediately.
The narrow conclusion is consequential: check whether a live service depends on an affected PHP runtime, apply the vendor or upstream fix, and treat isolation as temporary. The incident shows that AI can help uncover hard-to-find bugs. It does not show that all offline websites were compromised or that AI replaces the people who verify, disclose, patch and decide when a service can return.
FAQs
What is PHP2Shell?
PHP2Shell is the name Wiz gave to an exploit script combining vulnerabilities in PHP. Wiz says it demonstrated unauthenticated remote code execution against representative WordPress, Drupal and Joomla containers. The underlying flaws are in PHP, not a particular content-management system, so this is not a WordPress-only incident. CERT-EU identifies the PHP security updates as the complete remediation.
Were the websites taken offline successfully hacked?
Not necessarily. Reuters documented precautionary outages alongside attempted exploitation activity. Amazon said it could not determine whether attempts against its honeypot would have worked against real sites. A shutdown shows that an operator judged exposure unacceptable; it does not prove that the operator's systems were compromised.
Did an AI discover the vulnerabilities on its own?
No such conclusion is supported. Calif and Wiz said AI tools assisted their researchers, and Wiz named Atlas, Gemini 4 Argon and GPT-5.6 Sol. The PHP Foundation says people made the decisions and worked on the patches. The evidence supports AI-assisted discovery, testing and validation within human-led teams.
What should PHP operators do first?
Identify every internet-facing service that serves requests with PHP, determine its branch, and apply the appropriate vendor or upstream update. CERT-EU lists 8.5.12, 8.4.27, 8.3.36 and 8.2.35 as fixed versions and advises organizations that cannot patch immediately to consider isolation. Temporary measures do not make an installation safe.
Why did this become a global web story so quickly?
PHP is a foundational runtime used by many hosting environments and applications. A flaw in the interpreter can affect different products at once, even when their application code is unrelated. Public technical details shorten the response window. That explains precautionary outages without requiring the unsupported claim that every PHP website was attacked.