> status: mostly fine. patch anyway.practical AI security · est. 2026

Brief · Builders

Critical PHP Security Release: Patch, Rebuild, Skip the PoCs

By alkizz ·

You'd think the whole internet broke at 10 AM ET. It didn't. W3Techs says PHP is used by 69.3% of websites whose server-side language it can detect. That's most of the web, not the internet, and running PHP isn't the same as being exploitable.

What's confirmed: PHP's Oct 21 release fixes CVE-2026-109618 (RCE via out-of-memory corruption) and CVE-2026-109619 (RCE via stale global state). Each is CVSS v4.0 9.2, with high attack complexity, and gives unauthenticated code execution as the web worker. PHP says any install serving web requests is at risk: FPM, mod_php, php-cgi, FrankenPHP, LiteSpeed. mod_php is worst, because the demonstrated exploit fires before any app code runs.

Confirmed: threat actors are already trying. Cloudflare says it "observed attack traffic targeting our customers" and saw attackers "mutating the attack payload" to get past its protections. watchTowr told The Record it is seeing exploitation attempts in the wild, some using PoC tools built with AI's help, and that "nobody has been successful to date." Precursor Security reports attempts against both CVEs. No successful compromise is confirmed, and neither CVE is in CISA's KEV catalog as of the Oct 21, 2026 feed (catalogVersion 2026.10.21), re-checked about 12:45 PM ET. Sansec says writeups and an exploit tool are public and expects mass exploitation. Don't wait for the first confirmed breach. Patch now. A researcher's pre-release demo, which Huntress says was posted and then pulled, still isn't tied to these CVEs by PHP.

Before release, the gap filled with unverified PoCs; at least one author said AI models helped write theirs. Now watchTowr says AI-vibecoded PoC tools are being fired at real servers. None has worked yet, according to watchTowr.

The irony: AI helped with the real bugs too. Researchers at Calif, Wiz and OpenAI found and weaponized the bugs with help from frontier AI models; Sansec says all three credit frontier models with making the exploits practical. Wiz says it used Gemini 4 Argon and GPT-5.6 Sol; Calif says no special harness or models that weren't generally available were used in the initial discovery, and that knowing the target drove the find.

A random PoC may be fake, hit a different bug, or be malware aimed at people who run random PoCs. The config "mitigation" some traced to a chatbot answer, turning off session upload progress, was partly right: PHP now lists it as a mitigation for the session paths of CVE-2026-109619 only, not a fix. A fluent guess can be partly right, which is exactly why it's a hint and not a fix.

What to do:

It's real and worth patching today. It's also a patch, not the end of the internet. Update, rebuild, keep the logs, go back to work.

Sources

← All briefs