AI Didn't Find a Bug. It Found a Bug Class.
*The coverage says 'AI found a big flaw in PHP.' The real story is three independent finds, one deliberately unnamed mechanism, and a clock that started October 21.*
On October 21, 2026, word landed of CVE-2026-109618 — "php2shell," a flaw in PHP reportedly old enough to have been sitting in the codebase through multiple presidential administrations, with a blast radius estimated at up to 70% of the web. Government and public sites in multiple countries took themselves offline as a precaution (Times of India, Washington Post). The detail that should have been the headline, and mostly wasn't: the flaw was found independently by AI systems at both Google and OpenAI (Forbes, shattered.io).
Two independent systems arriving at the same buried flaw is not luck, and treating it as a neat anecdote — "AI found a big bug" — misses what it proves. Human code review is bounded by attention. Nobody reads twenty-year-old request-parsing code across an entire ecosystem; the flaw survived precisely because no human was ever going to hold all of it in mind at once. A system that reads at machine scale can sweep the whole corpus, which means a bug invisible to every individual auditor can be structurally visible to the sweep. That's the actual news: php2shell isn't a one-off, it's the first confirmed specimen of a bug class that human review cannot catch by construction. The next decades-old flaw of the same shape is already sitting in some codebase, waiting for the same treatment.
Now read the coverage closely — shattered.io, WaPo, CNN — and count what's missing: the mechanism. Nobody names how php2shell actually works. I'd bet that's deliberate.
The patch diff is the roadmap, and nobody wants to be the outlet that drew the map.
That throttle is rational, and it's also the first large-scale test of whether coordinated disclosure survives an era where discovery is cheap. The disclosure norms we inherited were built when *finding* a bug was the expensive, slow part — the work of a skilled team over weeks. If AI has collapsed the cost of finding, the chokepoint moves to the diff itself, and the entire discipline has to re-derive itself around that inversion. Holding the mechanism back buys time. It doesn't change the underlying math.
We've run versions of this race before, and the rhymes are instructive. Heartbleed, 2014: patch and disclosure landed the same day, mass scanning began within hours, but the blast radius was bounded and major infrastructure patched fast. Log4Shell, 2021: weaponized within hours of going public, and the tail dragged on for *months* because the flaw lived in dependency trees and unwatched servers that nobody could name an owner for. Both races taught the same lesson — the big properties patch in days; the long tail is where exploitation actually lives. php2shell's fate gets decided in the same place, and PHP's long tail — shared hosting, municipal sites, CMS installs nobody has logged into since the Obama administration — is longer than log4j's.
Which sets up the part everyone usually argues about in the abstract and can now watch in the specific: does AI arm attackers as fast as it arms defenders? The clock started October 21. Mass-patching the PHP long tail takes weeks. Weaponizing a public diff takes minutes. The first in-the-wild exploitation report is the data point that settles it — and notice how favorable the conditions are to the defense. The polite side found it first. Disclosure was coordinated. The mechanism is being held back. If attackers win the race anyway, that's the first empirical signal the offense/defense asymmetry has flipped, and no one gets to call it a hypothetical anymore. If patching wins, hold the champagne: that's one data point under best-case conditions, and the next decades-old flaw might not be found by the polite side first.