PHP fixes AI-found flaws attacked before the patch

By VK

Published 22 Oct 2026, 10:41 UTC

Reporting from TechCrunch, OpenAI, WIRED

AI-assisted researchers found two critical PHP flaws. Fixes shipped October 21, but AWS saw an exploit attempt four days earlier.

What it means for founders

Inventory every PHP server you rely on. Think WordPress marketing sites, WooCommerce or Magento stores, phpMyAdmin and Laravel apps. WordPress has told site owners the fix lives on their servers, TechCrunch reported, so update or get your host to confirm it has. OpenAI published a version checker that runs as a Codex skill.
Check logs back to October 17. AWS saw the first attempt that day. The UK's NCSC says to look for signs of compromise and backdoors. If you find any, the FBI's Brett Leatherman says to isolate the server and keep its web and PHP logs, and because leaked memory can hold secrets, treat that app's passwords and keys as exposed.
Budget for patching in hours. For critical flaws in internet-facing systems, AWS argues 7 or 30 day patch windows no longer fit. Edge mitigations from Cloudflare, AWS, Fastly and Vercel buy time but do not replace the update.
Watch for the next wave. OpenAI says PHP is not the last of its finds: more serious bugs sit in its fix queue, and it expects the pace to quicken over the next few months.

The story

An attacker who never logs in can, on some server setups, take over a machine running PHP. Fixes arrived on Wednesday, October 21, with attackers already probing. The PHP Foundation announced versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, closing two flaws rated 9.2 out of 10. PHP's security advisories say no workaround fully protects a server, and only the 8.2 line and newer receive official patches.

How AI surfaced a decade-old bug

The trigger is a fatal error, usually running out of memory, that cuts a request short at a bad moment. PHP can then keep using leftover memory it should have discarded, opening the way to running code or to leaking memory that, Wiz says, can contain passwords and API keys. The most dangerous paths fire before any site code runs, so clean WordPress or Magento code offers no protection.

According to OpenAI, security firm Calif used GPT-5.6 Sol under its Patch the Planet program to find the first flaw, reported on August 11. The PHP Foundation says OpenAI reported the same root cause that day. Google-owned Wiz reported more variants on August 24 with help from Google's Gemini 4 Argon, and argues frontier models squeezed months of elite exploit work into about a week.

A disclosure that sprang leaks

The CVE numbers began circulating on Tuesday, October 20, so PHP pulled the release forward by a day. According to the foundation, Calif briefly posted a demo video by mistake, and Chainguard shipped patched images more than six hours ahead of the agreed time. WIRED reported that Cloudflare saw early attacks before the release and, citing local reports and Reuters, that some government sites in Finland and France went dark as a precaution.

On Wednesday evening, AWS said one of its honeypots had been hit by an exploit attempt on Saturday, October 17, from an anonymizing network, and that it cannot say how the attacker got the details. "PHP2Shell is now being broadly exploited in the wild," Amazon security chief CJ Moses wrote, as WIRED reported. OpenAI maintains frontier AI found the flaws before attackers did, and pledged a $1 billion Daybreak subsidy for thinly staffed defenders.

What we don't know yet

How the October 17 attacker obtained working exploit details while the bugs were under embargo.
How many servers were breached before patching, and how many still run PHP versions that get no official fix.