Story
Oct 22, 2026
AI Found PHP2Shell First, but the Patch Race Is Already Underway
Researchers and AI companies cast PHP2Shell as proof that frontier models can find Internet-scale flaws before criminals do; security agencies and infrastructure operators see the same breakthrough as a warning that defenders now have only hours to react.
PHP2Shell did not emerge from nowhere. Two years before this week’s alarm, a researcher identified a flaw with the same root cause and said the PHP Foundation had been notified, yet no fix was deployed. The foundation later called that lapse a “structural failure” tied to its decision-making process and lack of resources.1
By mid-August, researchers at Calif and Google-owned Wiz had uncovered two critical PHP vulnerabilities using AI tools, including OpenAI’s GPT-5.6 Sol. The bugs, each rated 9.2 out of 10, could let an unauthenticated intruder execute code remotely on a PHP server — a daunting prospect for a language used by roughly 70% of websites with known server-side programming.1
The discovery initially looked like a defensive win. OpenAI said, “To our knowledge, frontier AI found these vulnerabilities before attackers did,” while warning that the same capabilities could help attackers quickly build offensive tools.2 That duality became the story’s central tension: AI had exposed a long-hidden weakness, but disclosure and patches could also give attackers a blueprint.
On Tuesday, Calif accidentally posted a proof-of-concept exploit video, accelerating a release that had been planned for Thursday. The PHP Foundation pushed patches out 24 hours early on Wednesday. By then, code claiming to exploit PHP2Shell had appeared on GitHub, although the foundation said it had seen no evidence of a working public exploit and described circulating generated code as incomplete or non-working.1
Governments treated the risk as immediate. The NSA urged organizations to patch at once or take systems offline; some European public websites went dark while operators worked. Its warning was blunt: these flaws can offer “a direct path to the server’s file system or database.”2
The race may already have started. Amazon said it detected an exploitation attempt four days before public disclosure and that the vulnerabilities were “now being broadly exploited in the wild.”2 Commercial tools rushed into the response too: Aravind Srinivas said users could receive credits to check PHP exposure, while promoting a tool that prioritizes fixes.
3
The episode leaves a harder lesson than simply patching PHP: AI may give defenders first sight of the next crisis, but it is also shrinking the interval before everyone else catches up.