Story
Oct 27, 2026
AI Found a Web-Wide PHP Flaw—Then the Internet Had Hours to React
AI helped researchers uncover a decades-old PHP vulnerability before attackers could capitalize fully. But leaked details compressed the response window, driving governments and companies to patch urgently or pull services offline.
In August, researchers at Calif and Google’s Wiz used advanced AI models to uncover related flaws in PHP, the decades-old language estimated to underpin roughly 70% of websites. The bugs could enable unauthenticated remote code execution—effectively giving an attacker command-line control of an unpatched server.1
For defenders, the discovery was a narrow win. Wiz researcher Sagi Tzadik initially doubted the AI’s finding because “models tend to hallucinate,” but the exploit worked. He argued that AI is accelerating vulnerability discovery: “months of traditional research condensed to, like, days.”2 Yet the same acceleration sharpened the danger once the findings escaped containment.
By mid-October, details began circulating before the planned patch release. A proof-of-concept video was briefly posted, the PHP Foundation brought its release forward, and attackers began probing. Cloudflare recorded thousands of attempts before disclosure, then tens of thousands within 45 minutes, hundreds of thousands within two hours and eventually millions.3
That prompted an extraordinary precautionary response on Wednesday and Thursday. Governments and businesses from Finland to Japan took sites offline while installing fixes; Britain said 32 public services were temporarily unavailable. For operators of legacy systems, the choice was stark: patch immediately or disconnect. Dutch cyber chief Matthijs van Amelsfort reduced the remedy to three words: “patch, patch, and patch.”4
The public-facing risk was more limited than the alarm suggested: merely visiting a vulnerable PHP site did not infect a user’s device. But a compromised server could expose data, alter sites or become a launchpad into wider networks. Early assessments found mostly scanning and failed exploits, while a later X repost citing honeypot activity said exploitation had been observed in the wild, with attackers issuing shell commands to fetch payloads.
5
The episode stopped short of a web-wide catastrophe because a patch was ready. It also exposed the new AI-era imbalance: defenders may find more flaws first, but every leak turns coordinated disclosure into a countdown.