PHP runs beneath websites built with WordPress, Joomla, Drupal and Wikipedia’s MediaWiki. Wiz Research says the language powers about 70% of websites worldwide. The PHP2Shell findings concern PHP’s core engine, so they sit below individual application bugs. The reported flaws date back to older releases, including PHP 5.4.
AI-assisted research helped turn those longstanding weaknesses into working exploits.
Wiz and OpenAI found vulnerabilities in core PHP
Wiz Research reported multiple critical vulnerabilities in PHP itself. OpenAI says its Patch the Planet initiative identified a critical flaw using GPT-5.6 Sol, while Wiz independently discovered and reported both vulnerabilities. The accounts describe parallel AI-assisted findings, not a claim that one organization reproduced the other’s work. Wiz’s PHP2Shell report and OpenAI’s coordinated disclosure detail the research.
The work used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol. The reported flaws include unauthenticated remote code execution, which can let an attacker run code on a remote system without logging in, and memory disclosure. Because the vulnerabilities are in PHP’s core engine, their risk does not depend on a vulnerability in WordPress, Drupal or another application running on top of PHP.
That shared layer is why the finding matters beyond any one product. PHP underpins many unrelated sites, but the language’s reported 70% share does not mean 70% of websites are vulnerable or compromised. It describes the breadth of PHP’s footprint, not a verified count of exposed servers.
PHP2Shell vulnerabilities carry a reported 9.2 score
The flaws were assigned CVE-2026-109618 and CVE-2026-109619. EdgeScan reports a CVSS v4.0 score of 9.2 for the vulnerabilities. CVSS scores communicate assessed severity; the score alone does not establish how many systems are exposed or whether a particular deployment can be exploited.
The affected history extends to PHP 5.4. Wiz reports that every PHP version released in the last decade is affected by at least one of the vulnerabilities. PHP released patches for multiple memory-safety flaws in the core engine on October 21, 2026. That closes the software-fix part of the story, but older installations and dependent applications can make identifying and applying the right update a slower operational task.
Successful exploitation could let attackers take control of affected servers, steal sensitive data or deface websites. Memory disclosure presents a distinct exposure: information held in a vulnerable process may be revealed. The available reporting describes these classes of impact, but does not establish that every affected installation has the same exposure or exploit path.
AI shortened the path from flaw to exploit
The vulnerabilities predated their discovery. Wiz says frontier AI models reduced the effort required to move from a longstanding flaw toward a practical exploit, with Gemini 4 Argon making an important contribution to finding a vulnerability. OpenAI’s separate finding adds evidence that AI can assist vulnerability discovery in a widely used software component.
The reported technical step was chaining: combining weaknesses or behaviors so one makes the next easier to exploit. EdgeScan says models helped researchers adapt exploits to specific targets, with working examples across several applications produced in hours. That is more consequential than simply generating code. It suggests AI can help researchers translate a core-runtime weakness into a test tailored to software that depends on the runtime.
OpenAI says the work demonstrates AI’s potential to uncover vulnerabilities with internet-scale impact and notes that the same capabilities could help attackers develop offensive tools and exploits. That is a risk assessment, not evidence that attackers used these models against PHP systems. The confirmed finding is narrower: AI-assisted research helped identify and test flaws whose potential reach is broad.
Faster discovery moves the bottleneck to remediation
The defensible forecast is not that existing security controls have become useless. It is that faster vulnerability discovery can compress the time defenders have to understand a flaw, identify affected systems and deploy a fix. Signature-based tools can detect known patterns, and manual testing can find serious defects. Neither guarantees that an organization will inventory every runtime version or update every dependent service quickly enough.
That creates a practical mismatch. AI can lower the cost of examining code and adapting tests, while remediation still depends on asset records, software ownership, maintenance windows and compatibility work. A faster exploit-development cycle makes those operational steps more important, not less. If a team cannot identify where an old PHP runtime is running, a patch announcement does not reduce its exposure.
The second-order effect is a shift in what security budgets buy. Continuous code analysis and automated testing could let teams examine more software more often, reserving scarce specialist time for validating findings and coordinating fixes. That gives AI-assisted defensive tools a plausible route to greater spending. The value would come from shortening the time between discovery and repair, not from the AI label itself.
My forecast is that within 24 months, security vendors offering continuous AI-assisted vulnerability discovery and testing will gain market share relative to products centered on periodic testing alone. The prediction depends on those tools producing useful findings that teams can validate and remediate without overwhelming them with false positives. It would be wrong if manual testing and signature-driven products retain their market share without adding comparable automated discovery and response, or if buyers find that AI tools fail to reduce remediation time.
The PHP case shows where the pressure lands: the speed of finding a weakness can rise faster than the speed of patching a distributed estate. AI does not remove the need for human judgment. It makes accurate inventories, clear ownership and coordinated updates more consequential because they determine whether a faster discovery process produces faster protection.
PHP operators face an inventory-and-patching race
The immediate task for organizations is to establish which PHP versions they run and where those runtimes sit in their infrastructure. The reported affected history reaches PHP 5.4, but the available facts do not quantify how many sites remain exposed. Actual risk therefore depends on deployment details and remediation status.
A patch must travel through real systems: administrators need to locate the runtime, confirm the relevant update and account for applications that depend on it. Where software is inherited or ownership is split across teams, that work can take longer than publishing a fix. The operational gap is not a flaw in the patch; it is the distance between a released correction and its installation across a complex environment.
If exploited, the reported vulnerabilities could give attackers control of affected servers, expose sensitive data or enable website defacement. Those outcomes make timely remediation material to both service control and data integrity. For PHP owners, the security test is whether they can find and update exposed runtimes before exploit development outruns their repair process.
PHP2Shell turns latent risk into a measurable defense test
PHP’s broad use made its core engine an important security boundary before these findings. PHP2Shell gives that risk specific identifiers, reported impacts and a patch date: CVE-2026-109618 and CVE-2026-109619, unauthenticated remote code execution and memory disclosure, and fixes released October 21, 2026. Wiz’s report documents the core-engine findings, while OpenAI’s disclosure records the separate AI-assisted discovery.
The evidence supports a change in discovery capability, not the end of human research or conventional security controls. The lasting test is whether defenders can turn faster discovery into faster inventory, validation and repair. That is the practical measure of whether AI-assisted security improves protection rather than merely accelerating the next exploit cycle.
