PHP Patches Two Critical RCE Bugs Found With GPT-5.6 Sol: the Flaw Chain Called PHP2Shell Reaches Back to 2004

PHP released security updates on Wednesday for two critical memory-management vulnerabilities that, when chained together, allow an unauthenticated attacker to run arbitrary code on a web server. The PHP development team urged every operator to upgrade immediately, warning that the flaws affect essentially all web service deployments built on the language.

Security researchers at watchTowr, whose analysis underpins the disclosure, have dubbed the exploit chain PHP2Shell. According to reporting by Ars Technica, the memory bugs reach back to every PHP release since PHP 5 shipped in 2004, which means two decades of deployed code is theoretically exposed until patched. The Wall Street Journal has estimated that as much as 70 percent of websites run on PHP in some form, from WordPress blogs to large enterprise applications.

The disclosure has an unusual twist: the bugs were surfaced in August with the help of OpenAI's GPT-5.6 Sol model, according to the WSJ report and the PHP project's advisory. OpenAI followed hours later with a statement of its own, saying the PHP vulnerability was one of "multiple significant vulnerabilities" discovered with the model and that it expects "similar findings will increase sharply."

Cloud infrastructure providers moved faster than the public advisory cycle. Cloudflare said it deployed default WAF mitigations for all customers, free and paid, roughly nine hours before the vulnerabilities were publicly disclosed, giving passive protection to sites whose owners had not yet patched.

Unauthenticated remote code execution is the most severe class of web vulnerability because it requires no credentials and no user interaction: an attacker can simply send a crafted request to a vulnerable server. Combined with the ubiquity of PHP, that is why the fixes are being treated as emergency patching rather than routine maintenance.

The episode is also a preview of how vulnerability discovery is changing. A frontier model finding a twenty-year-old class of memory bug shifts the economics of offensive research, and OpenAI's own framing — that such findings will "increase sharply" — suggests model-driven security work is a coming wave rather than a one-off.

For defenders, the practical takeaway is unchanged: patch now, and treat vendor statements about AI-assisted discovery, including those from OpenAI and Cloudflare, as self-reported claims. What is new is the clock speed. The interval between a model-assisted find, a CDN-level mitigation and a public advisory has compressed to a single day — and the next PHP2Shell may not come with nine hours of warning.