Story
Oct 27, 2026
AI Found PHP’s Hidden Flaws, but a Leak Turned the Fix Into a Race
Researchers and cyber agencies see AI as a necessary advantage for finding buried flaws before criminals do, while the PHP scramble showed that faster discovery means little if disclosure and patching cannot keep pace. For website operators, the lesson was less theoretical: patch quickly, or temporarily go dark.
In early August, Wiz researcher Sagi Tzadik asked an AI agent to test whether a weakness it had found in WordPress could translate to Drupal. The agent failed at that task—but surfaced a PHP flaw that had apparently sat undetected for more than a decade. Tzadik initially doubted the result: “Models tend to hallucinate.” The exploit worked nonetheless.1
A separate team at Calif, working with OpenAI, found a related vulnerability. The discoveries gave defenders an unusual head start. Wiz, Calif and PHP maintainers formed an informal group of roughly 15 researchers, sharing findings and building a patch over weeks in what Wiz CTO Tomer Schwartz called a process “just based on trust.”1 Their optimism rested on a hard reality: AI can compress months of security research into hours or days, but it can also hand attackers the same acceleration.
That danger became immediate in October. PHP underpins an estimated 70% of websites, and the flaws could let an unauthenticated attacker run commands on vulnerable servers. The issue was in PHP’s core, not an individual site’s code—meaning an unpatched small business page could present the same underlying exposure as a major institution.2
Then the coordinated-disclosure plan frayed. Details began circulating before the scheduled release; Calif accidentally posted a demonstration video, and Chainguard was accused by the PHP Foundation of breaking the embargo. Governments and companies took sites offline rather than risk intrusion. Cloudflare recorded thousands of attempted exploits before the announcement, then tens of thousands within 45 minutes, hundreds of thousands within two hours, and eventually millions.3
Authorities stressed that those were attempts, not proof of mass compromise. Germany’s cyber agency said it had not seen successful exploitation, while the Dutch agency’s prescription was blunt: “patch, patch, and patch.”3 For ordinary users, visiting a vulnerable site does not itself infect a device; the immediate burden falls on operators and hosts.4
The PHP episode is not isolated. OpenAI president Greg Brockman separately disclosed work on a remote-code-execution flaw in the widely deployed Brotli compression library, calling it an example of how difficult downstream patch coordination can be.
5 The emerging contest is clear: AI may help find the internet’s oldest cracks first, but defenders must close them before disclosure turns discovery into a map for attackers.