The PHP Bugs Were Twenty Years Old. The Wall Protecting Them Just Fell.

Two holes in PHP sat there for more than twenty years, present in every version since PHP 5 shipped in 2004, in a language that by Wiz's estimate underpins roughly 70 percent of the websites on earth. They were not safe because they were subtle. They were safe because exploiting them, in the security firm's own account, required an exceptionally complex chain of steps at the edge of what human researchers could practically achieve. That is not a defense. That is a load-bearing wall made of human limitation, and this week AI demolished it.

Security firm Calif, working in OpenAI's "Patch the Planet" initiative, used GPT-5.6 Sol to find one flaw; Google's Wiz used its Atlas agent alongside Gemini 4 Argon and GPT-5.6 Sol to find more, TechCrunch. Fine — defenders with good tools found old bugs. But Wiz's warning is what actually changed: before AI, these flaws were too hard for hackers to exploit, and now LLMs can adapt an exploit to different PHP versions, configurations, and memory layouts. The scarce resource in offensive security was never bugs; the backlog of known-but-impractical flaws across the legacy stack is enormous. The scarce resource was the expert labor to weaponize them. That scarcity just ended.

OpenAI says frontier AI found these vulnerabilities before attackers did. Amazon's threat intelligence team caught an exploit attempt on October 17th through a monitoring system built to bait hackers, The Verge — four days before public disclosure — and Amazon now says the flaws are being broadly exploited in the wild, Ars Technica. Cloudflare's CEO counted thousands of exploit attempts in the ten hours before announcement, hundreds of thousands in the two hours after, and millions by that night,. So the defender's head start was negative four days, and the attacker's scaling curve after disclosure is measured in model calls, not researcher-weeks. The NSA's advice — patch immediately or pull the system offline — is correct, and it is also a confession: once the difficulty barrier falls, the defenses that advice leaves are update speed and the power switch.

The bill is not the patch. Patching is cheap for anyone with a functioning ops team; the Internet Archive was secured and back online overnight. The bill lands on everything that cannot be patched. PHP 8.1 and older are no longer supported, and the recommended fix is an upgrade to 8.5.12 — a migration, not an update, for code whose maintainers left years ago. The Verge reports that World Quest and Save My Demonic World, two games running for over a decade on older tech their developers say cannot be updated, now face permanent shutdown. Some Dutch and Finnish government agencies took sites down as a precaution. 4chan and DeviantArt went dark temporarily. Some unmaintained corners of the web will simply never come back, and the internet archive of human culture gets a little thinner each time this happens.

AI found the bugs, coordinated disclosure produced patches before announcement, major providers were warned in advance, and the PHP Foundation says there has been no successful remote code execution to its knowledge. If you ran the experiment a hundred times, you would rather have AI on the defenders' side than not. I concede all of it, and it changes nothing structural. The attackers' conversion cost from "obscure bug" to "mass exploit" has collapsed permanently and symmetrically — Ars Technica notes that the cyber capabilities of this year's frontier models were strong enough that the US government temporarily limited their rollout. Defenders must now win a race that starts hours or days in the hole, against an adversary for whom producing a working exploit variant is no longer expert labor at all but a task the model performs. Meanwhile the patch treadmill — the one thing standing between disclosure and compromise — is a machine that a large, unmaintained, unpatchable fraction of the web cannot run at any speed.

So what does this do when it is real, repeated, at scale? Every ancient, ubiquitous, under-maintained layer of the stack — and the internet is built from them, PHP merely being the largest — now has an AI-assisted path from latent flaw to working exploit. The next PHP2Shell — the next bug with this one's reach from language runtime to remote shell — will not be in PHP. It will be in some library, some parser, some protocol implementation whose last commit predates the smartphone, and it will be found the same way, disclosed the same way, and exploited within the same hours.

If difficulty can no longer protect the long tail, then maintenance has to — and maintenance is a budget line, not a capability. We spent twenty years treating the unmaintained web as harmless legacy. This week is what the invoice looks like. Either someone funds the upkeep of the layers everything else stands on, or we keep paying in permanent outages, lost archives, and dark sites, one disclosure cycle at a time.