Frontier OpenAI Model Exposes Stunning Exploit That Leaves Millions of Web Servers Open to Attack
Security researchers used an OpenAI model to find two serious security flaws in PHP, the software that runs nearly 70 percent of all websites on the internet.
The flaws are so serious that all websites using PHP — millions upon millions of them — have been advised to take their servers offline right away and install the latest fix, which was released on Wednesday. After the news came out, large parts of the internet went offline. Websites that were unavailable for a time included Wikipedia, Steam, and even government websites.
However, many organizations may not have enough time, money, or staff to install the fix quickly. This could leave their systems open to attack until they do. According to the New York Times, the public information about the flaw could help attackers work out how to exploit it, take control of affected servers, and steal information. The National Security Agency advised that if website owners cannot install the fix, the safest choice is drastic: keep their websites offline and do not bring them back online.
The way the flaws were found was almost as remarkable as the scale of the threat. Researchers at cybersecurity company Calif discovered them using OpenAI's GPT-5.6 Sol model. They did this as part of a project with the AI company. Calif reported the flaws to the PHP Foundation on August 11.
Wiz, a security company owned by Google, found the flaws independently and filed its report on August 24. Wiz called them some of the “most severe vulnerabilities discovered in PHP, and arguably in software history.”
The exploits have also raised concerns about OpenAI's role in finding the flaws.
The flaws had existed in PHP versions going back to 2004. But through OpenAI's “Patch the Planet” program, Calif used the company's flagship GPT-5.6 Sol model to find them in just a few days. In a way, these exploits were uncovered because today's AI models are so powerful. OpenAI also encouraged researchers to find such flaws through its programs.
The way the flaws were made public has added to the confusion. OpenAI followed standard practices for responsibly reporting security problems, but some details about the attack leaked one day before the official announcement was scheduled.
Even if nothing had leaked, simply knowing that these flaws existed could still have caused enormous damage. AI played a large part in discovering them. This was the next step in the growing debate over AI companies' claims that their AI agents can carry out powerful hacking attacks. A headline in The Atlantic declared that the “AI-Hacking Apocalypse Began Last Night.” OpenAI will present the discovery as clear proof of what its AI can do, while the rest of the world deals with the damage to the internet.