AI has just revealed a flaw threatening the Internet

The flaw in the web’s plumbing

PHP2Shell is not another bug in a WordPress plugin, a theme, or a single web application. It is a pair of vulnerabilities in PHP itself, the runtime that processes requests before much of the application code even has a chance to run. On October 21, 2026, the PHP development team released PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, describing the update as a security release for two critical vulnerabilities, CVE-2026-109618 and CVE-2026-109619, that can allow unauthenticated remote code execution [1].

That phrase is the alarm bell: “unauthenticated” means an attacker does not first need a password, account, or stolen session; “remote code execution” means the attacker may be able to make the server run code of the attacker’s choosing. Tenable’s emergency FAQ records the two flaws as PHP remote code execution issues rated CVSSv4 9.2, placing them in the critical category that security teams usually reserve for patch-now events [6].

The vulnerabilities are collectively known as PHP2Shell, a name that captures the fear clearly enough: turning a PHP request into shell-level control. Cloudflare says the flaws affect web servers running PHP 5.4 or later and can allow denial of service, data exposure and remote code execution when an application processes specially crafted HTTP traffic [4]. PHP’s own release announcement is narrower in its practical patch guidance, naming current supported releases and telling users that updating is the only safe choice [1].

Why AI is central to this story

The headline is not just that PHP had serious memory-safety bugs. The disturbing part is who found them, how quickly they were refined, and how fast attackers could adapt once details began to leak.

OpenAI says Calif identified a critical PHP flaw using GPT-5.6 Sol as part of the Patch the Planet initiative, while Wiz independently discovered and reported both vulnerabilities [2]. Wiz says it used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol throughout its research, and specifically argues that AI helped move complex memory-corruption work from elite, slow, manual research toward faster, repeatable analysis [3].

That matters because these were not simple “copy this parameter into a command” bugs. Wiz describes the flaws as memory-safety issues rooted in how PHP handles fatal errors during request processing, especially out-of-memory conditions that can leave corrupted or stale internal state behind for later requests [3]. In practical terms, the runtime can trip, fail to clean up properly, and then a later request may interact with leftovers from the previous failure.

Pre-AI, adapting an exploit like that across different PHP versions, server APIs, memory layouts and applications required rare specialist skill. Wiz’s warning is that LLMs can help analyze crashes, compare failures, iterate payloads and adapt exploit chains across deployments, reducing the cost of turning a vulnerability into a usable attack [3]. The same capabilities that help defenders find buried flaws also help attackers replay the logic at scale.

What researchers demonstrated

The most worrying demonstrations were not theoretical. Wiz says it combined the vulnerabilities into a single exploit script called php2shell and demonstrated unauthenticated remote code execution against stock container deployments of WordPress, Drupal and Joomla as representative targets [3]. Wiz also says the underlying issue is in PHP itself, not in those applications, so exploitation does not depend on a CMS-specific coding mistake [3].

The impact also extends beyond getting a shell. Wiz reports that attackers could disclose sensitive data from PHP worker memory, including passwords, tokens and API keys, without necessarily achieving full remote code execution [3]. That is the nightmare middle ground for responders: a server might not show obvious signs of takeover, yet secrets passing through a long-running worker process could already have leaked.

Cloudflare’s public explanation gives the higher-level version: one flaw involves out-of-memory handling interrupting PHP while internal request data is only partly initialized, while the other involves request-owned state surviving after backing memory has been freed [4]. Those details are enough to understand the defensive urgency without publishing a recipe for exploitation.

The patch race became a public incident

The coordinated release did not unfold quietly. PHP’s announcement landed at 14:10 UTC on October 21, telling all PHP 8.5 users to upgrade immediately and stating that no complete mitigation exists besides updating [1]. Tenable reports that the PHP security team had issued an advance notice on October 20 for an October 21 security release, after CVE identifiers and discussion began circulating publicly [6].

Cloudflare says it deployed web application firewall protections from early October and enabled protections for free and paid WAF customers by default starting at 03:02 UTC on October 21, roughly ahead of public disclosure [4]. That kind of pre-positioning shows how seriously infrastructure providers treated the bug: when the vulnerable component is PHP itself, protection has to happen at many layers at once, from runtime packages to hosting platforms to edge firewalls.

AWS then added a more uncomfortable data point. Amazon Threat Intelligence says its MadPot honeypot network recorded a PHP2Shell exploitation attempt on October 17, four days before public disclosure of CVE-2026-109618 and CVE-2026-109619 [5]. AWS also says proof-of-concept exploits appeared on October 21 before the embargo lifted, with additional attempts from anonymized infrastructure, and that broad exploitation attempts were seen after patch release [5].

That changes the response clock. If exploitation may have started before disclosure, administrators cannot simply patch and move on; they must review logs and assume some probing may predate the official advisory.

Why websites went dark on purpose

One of the strangest visible effects was not a wave of defacements, but precautionary silence. UptimeRobot reports that governments, hospitals, media sites and hosting companies intentionally took websites offline on October 21 while waiting for a fix, with examples including Dutch public services, French prefecture sites and news services [7].

The monitoring data shows why this felt global. UptimeRobot says downtime alerts on October 21 rose about elevenfold over a normal day at peak, with Europe hit hardest; monitors from European locations saw alerts rise about twentyfold, and several European country domains showed sharp jumps in unreachable sites [7]. This was not necessarily evidence that all those sites had been hacked. It was evidence that operators judged the window between disclosure and patching to be dangerous enough to justify downtime.

That is a rational response to a runtime-level flaw. If the application layer cannot fully protect itself, and if the patch is minutes or hours away, temporarily shutting exposed PHP services may be less damaging than leaving them online as exploit code mutates.

What should happen now

The immediate answer is boring and urgent: upgrade PHP to a fixed build, restart workers, and verify that hosting providers have patched managed environments. PHP says updating is the only safe choice for the supported branches it released [1]. Cloudflare also stresses that changing website code or plugins alone does not fix the issue, because the flaws live in the PHP runtime [4].

The second answer is incident response. AWS recommends investigating from at least October 17, looking for suspicious POST requests containing PHP_SESSION_UPLOAD_PROGRESS, reviewing web server logs, and checking for unexpected process execution or file modifications on web application servers [5]. Wiz recommends paying attention to unusual multipart requests followed by memory-limit errors or repeated HTTP 500 responses, malformed or unexpectedly large PHPSESSID cookies, and unexplained PHP worker churn [3].

The strategic answer is more uncomfortable. AI did not create PHP2Shell, but it made the hidden weakness legible. It helped researchers find and validate a deep flaw, and it lowers the barrier for adversaries to adapt similar bug chains once patches, diffs, or partial details are visible. The web’s backbone has not collapsed, but it has been reminded that old code paths, obscure error handling and “too hard to exploit” assumptions are exactly where modern AI systems like to dig.

Before Skynet wakes up, the web’s backbone needs more than a debug session. It needs faster patch pipelines, better memory-safety investment, and defenders using the same AI acceleration that attackers now have.

Sources from the last 72 hours

1. [1]php.announce: [ANNOUNCE] PHP 8.5.12 ReleasedOct 21, 2026, 4:10 PM
2. [2]Patch the Planet: global response coordination as agents uncover critical vulnerabilitiesOct 21, 2026, 2:00 AM
3. [3]PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the InternetOct 21, 2026, 2:00 AM
4. [4]Cloudflare deploys protections against php2shell vulnerabilityOct 21, 2026, 2:00 AM
5. [5]php2shell (CVE-2026-109618, CVE-2026-109619): Critical PHP RCE FAQOct 21, 2026, 2:00 AM
6. [6]Inside PHP2Shell: Why Websites Took Themselves OfflineOct 22, 2026, 2:00 AM
7. [7]Amazon Threat Intelligence identifies PHP2Shell exploitation prior to public disclosure in MadPotOct 21, 2026, 2:00 AM

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