The PHP flaw credited to AI had been spotted as early as 2024

Public websites pulled offline as a precaution

CERT-EU, the cyber response team for the EU institutions, issued a rare piece of advice this week: if you can't update PHP right away, consider taking your exposed applications off the public network. It's like telling a shopkeeper to close up until the lock is changed. According to The Register, Finnish public services and several Dutch sites, including the Parliament's and the data regulator's, were taken offline as a precaution.

PHP is the language behind a large share of the web. On October 21, a day earlier than planned because of a leak, the project shipped versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. They fix two critical flaws, scored 9.2 out of 10, that can let an unauthenticated attacker run code on a server under certain conditions. The project warns that no workaround replaces the update.

"AI found the flaw"

The story arrived with the patches. OpenAI ranks the affair among the most serious security problems of recent years. It credits the firm Calif, which found the flaw using GPT-5.6 Sol, OpenAI's model, as part of its Patch the Planet program. To its knowledge, OpenAI writes, AI found these flaws before attackers did.

Wiz, which found them independently, praises the "invaluable" contribution of Gemini 4 Argon, Google's model. In the same post, Raluca Ada Popa, who leads Gemini security at Google DeepMind, explains that the flaws sat unfixed for more than ten years because finding and exploiting them was too complex. The Decoder's headline says OpenAI's and Google's models found the flaw.

In that telling, AI is the truffle hound that digs up what nobody had sniffed out. The same day, the PHP Foundation published its own version.

The foundation's side of the story

In 2024, PHP went through its first external security audit in years. The foundation had secured the funding from the Sovereign Tech Agency, a German public agency. The audit was organized by OSTIF and carried out by the French firm Quarkslab.

While fixing one of the issues Quarkslab flagged, a PHP core developer under contract with the foundation spotted the first signs of a memory-handling problem. The same one that sits behind the first of the two flaws fixed this week. Exploitation was considered only in theory, no concrete way to achieve it was found, and other flaws with an obvious impact took priority.

The security team then drafted several competing fixes. Each had a price: breaking compatibility with existing code, opening a new denial-of-service risk, or forcing changes across a large number of third-party extensions, a coordination effort that looked impossible. The team decides by consensus, feared shipping the wrong choice, and the few developers who know the engine well enough were busy elsewhere. The work stalled and restarted for months.

Picture a homeowners' association with several quotes to repair a cracked beam, none without a downside, that insists on a unanimous vote. The foundation itself speaks of a "structural failure", caused both by how the team made decisions and by a lack of resources.

What AI changed

On July 30, 2026, Calif began auditing PHP. Its researcher, Ryan Austin, started from a flaw found in 2004 by Stefan Esser, then the leading name in PHP security, and asked the AI to hunt for descendants still alive in the code. The model served as a flashlight, but it was the researcher who knew which cellar to go down. His conclusion, on Calif's page: you always have to know your target, and no model digs up every bug from a simple prompt.

On August 11, Calif reported a remote code execution with the same root cause as the problem seen in 2024. OpenAI, for its part, reported the same cause that same day. Wiz followed on August 24. Calif even acknowledges this in its timeline: a related problem had been identified in 2024, with no proof it could be exploited.

That is what tipped everything. According to the foundation, those reports made the real impact obvious, and from August 11 the fix became the top priority of its entire security team, down to working weekends. AI didn't reveal an unknown. It supplied the demonstration that was missing to settle the question.

Chainguard's CEO, Dan Lorenc, told The Register he rediscovered the flaws in a single evening with a frontier model and a basic prompt. He had just learned from a researcher that something was brewing in PHP. Knowing where to look, again. And on the fixes, he is blunt: it really did take months to build one that works.

Finding bugs is no longer the problem

Reports now arrive faster than they can be handled. Since July 1, PHP's security team has published 15 CVEs, the public identifiers for flaws, and closed 106 reports that weren't vulnerabilities, were duplicates or couldn't be reproduced, according to a foundation post dated October 14. It's an emergency room where ambulances arrive faster than surgeons can be trained.

At curl, maintainer Daniel Stenberg says, as quoted by The Register, that minor flaws will now be handled in public, for lack of time to deal with them in private. PostgreSQL pulled about a dozen features from version 19, and one of the explanations put forward is the flood of defects surfaced by AI-assisted review.

The foundation draws a clear line. The tools find bugs. They can't pick the right fix for an old design choice in the engine, weigh what it breaks, or review the patch. That takes engineers who know all of PHP's code, and the foundation writes that there are too few of them.

Since spring it has been reorganizing: a dedicated ecosystem security team in May, two hires in July and September, and decisions that should be possible by majority vote. It has run PHP through Anthropic's Claude Mythos and validates reports with OpenAI's Daybreak. But every decision and every patch is still a human's job.

One public funder named

That leaves the money. PHP runs public services, banks and a good part of corporate software. Yet its security rests, according to the foundation, on a small group of developers it pays for and some volunteer help. It asks companies, hosting providers and governments to pay in proportion to what they get out of it.

Across its whole post, one public funder is held up as an example: the German agency that paid for the 2024 audit. The foundation wants more, and "lasting" backing at that. Yet that audit is what put a developer on the trail. It took two years, and a demonstration obtained with the help of AI, for the trail to be followed to the end.