geek vibes.geek vibes.
Tecnología · 7 min · Oct 25, 2026

PHP2Shell: AI Found the Flaw Hitting 70% of the Web — Then Exploited It in 18 Hours

Two critical PHP flaws — the language behind 70% of the web — went from a leaked proof of concept to active exploitation in 18 hours. The same AI that helped find them is now used to patch them too. What it means for any agency managing client sites.

Ilustración editorial de un escudo de ciberseguridad agrietado junto a líneas de código y un reloj digital, en tonos navy, verde y cian, sobre fondo oscuro degradado.

On October 20, around 8pm UTC, cybersecurity firm Calif accidentally uploaded a video showing a working exploit for two critical PHP flaws — the language running behind WordPress, Drupal, Joomla, and Magento, and which, according to W3Techs, powers close to 70% of the world's websites whose server-side language is known. Eighteen hours later, at 2:19pm UTC on October 21, security firm watchTowr was already reporting active exploitation attempts against both flaws. At Geek Vibes we manage infrastructure on this exact stack for several clients, so this isn't a theoretical story: it's the week the window between "there's a patch" and "you're already being probed" shrank to hours — and AI was the accelerant on both sides of the board, offense and defense.

The flaws: what PHP2Shell actually is

PHP2Shell is two unauthenticated remote code execution (RCE) vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619, both rated 9.2 on CVSS. The first is an out-of-memory flaw in the PHP runtime; the second is tied to how PHP handles session upload progress (session.upload_progress), which runs before PHP finishes setting up the request. Even without achieving code execution, an attacker can read worker process memory — where passwords, tokens, API keys, and certificates tend to live.

What's unsettling is that the underlying code has sat in PHP for more than a decade; Calif traced its lineage back to research published in 2004. It affects nearly every version since 5.4, and every way of deploying PHP: PHP-FPM, Apache mod_php (the highest-risk path, with a demonstrated single-request exploit), php-cgi, FrankenPHP, and LiteSpeed. A Shodan scan cited by researcher Henk Poley counted more than 1.9 million internet-exposed hosts running PHP; researcher Silas Cutler estimated over 1.3 million are potentially vulnerable.

The timeline that should worry any agency

  • October 17: Amazon's MadPot honeypot network had already logged an exploitation attempt — four days before public disclosure. Amazon still doesn't know how that attacker got exploit information so early.
  • October 20, ~8pm UTC: Calif accidentally publishes the video with its working proof of concept. The same day, a Hacker News post leaks the CVE numbers, forcing the planned October 22 release forward to October 21.
  • October 21, 3:02am UTC: Cloudflare turns on firewall rules by default for all customers as a stopgap, before an official patch even existed.
  • October 21, 7:41am UTC: Chainguard ships already-patched container images, six hours ahead of PHP's own release.
  • October 21, 2:00pm UTC: PHP releases the patched versions.
  • October 21, 2:19pm UTC: watchTowr confirms active exploitation attempts against both CVEs — 19 minutes after the patch went out.

Amazon says the October 21 wave of attacks suggests multiple independent actors, and now recommends that any incident investigation start from October 17, not from disclosure day.

AI playing on both sides of the board

On the offense side: Wiz Research, Calif, and OpenAI's own Codex Security team (using its GPT-5.6 Sol model through the "Patch the Planet" initiative) worked in parallel with PHP's security team and Google's Gemini team. Calif later credited independent researcher @lcfr_hax with first finding PHP2Shell and building the original working exploit.

On the defense side: the very same week, Amazon put its autonomous pentesting agent — now part of AWS Continuum — through a live test against 37 recently disclosed CVEs (not PHP2Shell specifically). The agent captured the flag through the intended vulnerability path end-to-end in 33 of 37 cases, 89.2%, with a median time-to-capture of about 4 hours, running on Claude Opus 5 with a 10-hour cap per sample. AWS is upfront about the limits: it's their own benchmark, not an independent one; the sample is small; training-data contamination isn't fully ruled out; and there were two misses where the agent crashed the target before it could read the flag.

The point isn't that one side "wins." It's that the same AI capability that turns a leaked flaw into a working exploit in hours is, literally, the same technology big vendors are already using to find and patch flaws autonomously. The gap that matters now is who deploys that speed first — and it's measured in hours, not news cycles.

How exposed is your site, really?

The patched versions are 8.5.12, 8.4.27, 8.3.36, and 8.2.35; source backports exist for 7.4, 8.0, and 8.1, which no longer get official project support. PHP 8.2 reaches end of life on December 31, 2026, so it isn't a permanent place to park.

One point that trips people up under pressure: updating WordPress, its plugins, or its theme does not touch the server's PHP install — they're completely independent layers. Configuration-level mitigations (adjusting memory_limit, post_max_size, or disabling session.upload_progress.enabled) reduce exposure, but they're not a substitute for the real patch.

There's also an uncomfortable detail about the process behind all this: a PHP core developer had already flagged the out-of-memory bug during a 2024 external audit, but the fix stalled for months because competing fix approaches each had their own drawbacks. The PHP Foundation itself later called that delay "a structural failure" of its decision-making process. In the meantime, companies including BambooHR and several Dutch government and education sites chose to take services offline as a precaution.

What this means for agencies and developers

  • Treat "there's a patch" as "you're already being probed." The jump from leaked proof-of-concept to confirmed active exploitation was 18 hours. A weekly or monthly patch cycle isn't a security policy anymore — it's a bet.
  • Verify the actual PHP version on the server, not just the CMS. An up-to-date WordPress running on a vulnerable PHP install is still a vulnerable site.
  • Decide today who has the authority to ship a hotfix without waiting for a maintenance window. If the answer involves a committee or a ticket that takes days, that's the process to fix first.
  • Look at autonomous security agents (AWS Continuum and similar) without the hype filter. If the offensive side already operates at machine speed, defending at human speed stops being enough.
  • Put in writing, in your service contracts, who owns runtime patching (PHP, Node, whatever the stack is) versus who manages the CMS layer on top. This incident showed that line confuses a lot of people exactly when it needs to be clearest.

If you genuinely don't know what PHP version the sites you manage for clients are running right now, that's the message to send today — not after you finish reading this. At Geek Vibes we help teams audit their stack and put real urgency behind patching when the clock is measured in hours. Email us at hello@geekvibes.agency if you want us to check your exposure.

Does any of this resonate?

If you're running pieces that don't quite connect, we'll diagnose it together in 30 min.

→ Schedule a diagnosis
ACTIONS