PHP Flaws Found by AI Spark a Race to Patch the Web

Emergency patches shipped on October 21, 2026 for two critical bugs in PHP, the programming language quietly running most of the websites you visit every day, and the way the PHP Flaws were found may matter as much as the fix. The vulnerabilities, nicknamed PHP2Shell by the researchers who disclosed them, can let an attacker run code on a vulnerable server without logging in, according to Wiz Research. Frontier AI models did a meaningful share of the detective work, in what Wiz called an AI inflection point for security research.

The bugs are memory-safety defects in the core engine of PHP itself, not in any particular app or plugin. Roughly seventy percent of websites run on the language, including WordPress, Drupal, Joomla, and the software behind Wikipedia, so a flaw at this level reaches an enormous share of the web. One of the two issues lets an attacker corrupt a server's memory after an out-of-memory error and then run code; the other leaves references to freed memory that a later request can exploit. Both were rated 9.2 out of ten for severity, and vulnerable code stretches back more than a decade.

How AI helped find the PHP Flaws

Wiz said it used its own Atlas agent together with two frontier models, Gemini 4 Argon and GPT-5.6 Sol, and credited the models with making the hunt practical. "Frontier models were critical to making this process practical, enabling us to develop and adapt working exploits across multiple applications in hours," the company wrote in its disclosure. A single researcher using those models did in about a week what might once have taken a team of specialists months, according to the report.

The discovery was not a one-lab story. Security firm Calif independently found one of the flaws using a frontier OpenAI model through the company's Patch the Planet program, while Quarkslab also contributed to the coordinated disclosure, according to Cloudflare's tracking. OpenAI described the find as one of the most consequential security issues in recent memory and said it has committed a billion dollars in subsidized access to its Daybreak security tooling for under-resourced defenders. That framing points at the real shift here: AI is no longer just the thing being secured, it is doing the securing.

The race between disclosure and attackers

The timeline shows how fast these moments now move. Wiz first spotted the initial bug in mid-August and reported both flaws to the PHP project later that month, setting up a coordinated disclosure. But details leaked publicly a day ahead of the planned announcement, forcing the PHP team to accelerate its release schedule, and security firms reported exploit attempts, some landing before the fixes were even out. The patches arrived across all four supported PHP branches in a single coordinated release.

Governments treated the flaws as urgent. The NSA advised taking servers that cannot be patched offline, and the Netherlands' cybersecurity center pulled its own site offline as a precaution while it assessed exposure. The PHP project flagged one common server setup as the highest-risk configuration and demonstrated that the attack works before any website code runs, meaning no theme, plugin, or firewall rule can fully substitute for the update itself.

What it means for everyone else

If you run a website, the message is simple: identify your PHP runtime and update it now, following the vendor's advisory for your hosting setup. Note that updating WordPress or a CMS alone does not fix the underlying exposure, as the security firm Wordfence confirmed, because the flaw lives in the language beneath the site. Where patching cannot happen immediately, isolating or monitoring the affected server is the next-best move until the update lands.

For everyone who just browses the web, there is nothing to click and nothing to download, but the week is a useful reminder of how much of daily life sits on software you never see. Most breaches that follow a flaw like this one arrive not through the flaw itself but through password reuse and phishing in the weeks after, when attackers exploit the window between disclosure and patching. Turning on two-factor authentication and using unique passwords remains the cheapest defense against whatever comes next.

Zoom out, and the PHP Flaws story is really two stories. One is a routine, if severe, patching event: old code, critical bug, fast fix. The other is the new normal in which AI systems find the bugs, help write the exploits, and help defenders race the clock. This week both sides used the same tools, and the side that patched faster won. Read more of our coverage on the Gen Z security gap, leaner AI models, and OpenAI's health-tech bets. Full technical detail is in the Tesla News Wire report, the Pivot News breakdown, and the OpenTools overview.