OpenAI's ChatGPT Sol uncovered critical PHP vulnerabilities affecting 70% of websites

The headline matches the subject: ChatGPT Sol and PHP2Shell

OpenAI’s ChatGPT Sol story is no longer a theoretical debate about AI-assisted bug hunting. On October 21, 2026, OpenAI said that Calif identified a critical PHP flaw using GPT-5.6 Sol as part of its Patch the Planet initiative, while Wiz independently discovered and reported both vulnerabilities; PHP maintainers then led the fixes and release with help from OpenAI, Wiz, Cloudflare and others [1]. The vulnerabilities are now tracked as CVE-2026-109618 and CVE-2026-109619, and the emergency response matters because PHP sits underneath a vast share of the public web, including popular publishing, commerce and enterprise applications [3].

The shorthand “affecting 70% of websites” needs careful wording. The issue is not that every website is known to be exploitable. Rather, PHP is used by roughly 70% of websites whose server-side language is known, and these flaws are in the PHP runtime itself, below WordPress, Drupal, Joomla, Magento and other application layers [4]. That is why the story grew so quickly from an AI-assisted research win into a global patching event.

What was found

The PHP project published fixes for PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, warning that the two flaws can allow unauthenticated remote code execution and urging all PHP users to update immediately [3]. PHP’s advisory describes the first issue, CVE-2026-109618, as remote code execution through out-of-memory corruption: an allocation failure can interrupt PHP while request data is only partly initialized, after which request cleanup may treat invalid data as valid [3]. The second issue, CVE-2026-109619, is remote code execution through stale global state: request-owned internal state can survive after its backing memory is freed, and a later request handled by the same worker may use that stale state before initialization resets it [3].

That technical distinction is important. A normal website security review often focuses on application code, plugins, themes, input validation or access control. PHP2Shell sits deeper. Cloudflare emphasized that updating website code or plugins alone does not fix these vulnerabilities because the affected component is the PHP runtime [4]. HackerOne made the same operational point: these flaws affect the runtime, not a specific content management system, plugin or application [8].

How Sol changed the timeline

The most consequential part of the episode is not merely that AI helped find a bug. It is that AI helped revisit a bug class that humans had seen before but not fully converted into a patch-driving exploit case. Calif’s write-up traces the path from a 2020 PHP bug report involving out-of-memory behavior, to a 2024 audit that found a related issue without proving exploitation, to a 2026 Patch the Planet effort in which Calif began looking at PHP again [2]. Calif says that after starting the audit around July 30, researchers surfaced the first out-of-memory fatal issue by August 1, reported the first bug to PHP on August 11, shared a remote-code-execution proof of concept on August 14, and had PHP confirm that the PoC worked on August 15 [2].

That is the two-week transformation at the heart of the story: an old memory-handling concern became a working RCE demonstration fast enough to reshape the disclosure process. OpenAI frames the result as a “Defender’s Window,” saying frontier AI found the vulnerabilities before attackers did, allowing maintainers and providers to prepare fixes and protections ahead of public disclosure [1]. But the same OpenAI post warns that the capability cuts both ways: AI can uncover internet-scale flaws, while also helping attackers develop offensive tools and exploits more quickly [1].

Why PHP2Shell is so severe

Wiz’s technical analysis says the disclosed vulnerabilities are memory-safety issues arising from how PHP handles fatal errors while processing requests, especially out-of-memory errors [5]. Those failures can leave corrupted internal state behind after a request ends unexpectedly; a later request or worker can then reuse the stale state, causing memory disclosure or memory corruption [5]. Wiz says it combined the vulnerabilities into a single exploit script named php2shell and demonstrated unauthenticated remote code execution against WordPress, Drupal and Joomla deployments, while stressing that exploitation does not depend on a bug in those applications because the weakness is in PHP itself [5].

Calif reported broader testing as well, saying it confirmed pre-authentication RCE on multiple PHP applications and, in some cases, arbitrary PHP scripts [2]. The lesson is uncomfortable for defenders: even a simple PHP app can be exposed if the underlying runtime is vulnerable, and security controls implemented inside the application may never get a chance to run. Datadog’s reproduction reached the same conclusion for CVE-2026-109618: the malicious request did not reach the target application’s code before PHP executed the attacker’s command during cleanup of a failed request [7].

The impact is not limited to command execution. Wiz also said attackers could read sensitive data from PHP process memory, including passwords, tokens and API keys, even without first achieving full remote code execution [5]. That makes incident response more difficult, because teams cannot treat the absence of a visible webshell as proof that secrets were not exposed.

The emergency response

The PHP team moved quickly once the issue became public. Its October 21 advisory says all PHP installations serving web requests are at risk, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed, with risk highest for mod_php because a demonstrated unauthenticated exploit can trigger before application code runs [3]. For supported branches, PHP named the fixed versions and told users of Linux distributions or long-term-support vendors to check for updated packages, because vendors may backport fixes for older branches [3].

Cloudflare said it had been part of the coordinated mitigation effort and had deployed WAF protections ahead of public disclosure; it also said it observed attack traffic targeting customers and saw attackers mutate payloads in attempts to bypass protections [4]. The company stressed that WAF rules are defense-in-depth, not a replacement for patching, and told administrators to keep protections enabled while updating PHP itself [4]. Corelight, meanwhile, released open-source Zeek and Suricata detections so defenders could investigate suspicious HTTP behavior and prioritize patch review [9].

U.S. agencies also escalated the warning. The NSA said PHP2Shell could allow attackers to take control of affected web servers without prior access or authentication, and recommended that organizations inventory PHP-dependent services, apply vendor patches and take systems offline if immediate updating is not feasible [6]. Nextgov reported that the NSA, FBI and CISA urged organizations to fix the flaws, while noting that Palo Alto Networks’ Unit 42 had identified more than 22.6 million internet-exposed instances potentially vulnerable to the issues [8].

What administrators should do now

The safest instruction is also the least glamorous: patch PHP. The fixed upstream versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 [3]. Administrators should not assume that updating WordPress, Drupal, Joomla, Laravel, Magento or a plugin resolves the problem. The runtime must be updated, containers rebuilt, workers restarted or replaced, and managed hosting providers asked to confirm their patch status.

Teams should inventory every PHP runtime serving HTTP requests, including containers, sidecar services, legacy virtual hosts, control panels, vendor appliances and staging systems. Datadog warns not to rely only on the command-line PHP version, because the PHP interpreter serving a web application may differ from the CLI installation [7]. For older PHP branches, teams should check operating system and LTS vendor advisories, because the PHP project no longer normally updates PHP 8.1 and older, although it provided minimal source backports for PHP 7.4, 8.0 and 8.1 [3].

Temporary mitigations can reduce exposure to some attack paths, but the PHP project and security vendors repeatedly say they are not complete fixes [3][7]. Useful investigation signals include unusual multipart requests followed by memory-limit errors or repeated HTTP 500 responses, malformed or unexpectedly large PHPSESSID cookies, unexplained PHP worker crashes, and web server processes spawning shells or other unexpected child processes [5][7]. Absence of those signals does not prove safety; some exploitation paths may leave little application-level evidence because they occur below the application layer [5].

The bigger AI-security lesson

PHP2Shell is a turning point because it compresses three realities into one incident. First, critical internet infrastructure still depends on old C and C-style runtime assumptions where memory errors can hide for years. Second, open-source maintainers need coordinated help when a flaw affects infrastructure used by millions of sites. Third, frontier AI is changing the economics of exploit development.

OpenAI wants the story read as a defender success: AI-assisted researchers found the bug class, a coalition coordinated disclosure, and maintainers shipped patches before broad attacker exploitation was confirmed [1]. That reading is valid, but incomplete. The same tools that helped Calif, OpenAI and Wiz reason through PHP internals could eventually help less responsible actors compress their own timelines. The durable answer is not panic; it is faster inventory, better patch pipelines, layered detection and stronger support for the maintainers of the software the web quietly runs on.

Sources from the last 72 hours

1. [1]Patch the Planet: global response coordination as agents uncover critical vulnerabilitiesOct 21, 2026, 2:00 AM
2. [2]php2shell by Calif ResearchOct 21, 2026, 2:00 AM
3. [3]Cloudflare deploys protections against php2shell vulnerabilityOct 21, 2026, 2:00 AM
4. [4]PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the InternetOct 21, 2026, 2:00 AM
5. [5]PHP2Shell - Critical vulnerabilities in PHP (CVE-2026-109618, CVE-2026-109619): detection, analysis, and remediationOct 21, 2026, 2:00 AM
6. [6]US agencies urge fixes for programming language flaws putting millions of web systems at riskOct 21, 2026, 6:04 PM
7. [7]php2shell: Zeek and Suricata Detections for PHP RCEOct 21, 2026, 3:48 PM
8. [8]Security advisory: Critical remote code execution vulnerabilities, update nowOct 21, 2026, 2:00 AM
9. [9]NSA Advisory: Critical Patches Recommended for PHP VulnerabilitiesOct 21, 2026, 2:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.