OpenAI Says Its AI Models Helped Discover Critical PHP Flaws

Artificial intelligence helped discover critical security vulnerabilities in the programming language PHP, forcing widespread emergency patching and temporary website outages across the internet. According to OpenAI, the flaw, nicknamed PHP2SHELL, could allow attackers to take over servers running PHP, which powers a large share of the web. OpenAI reported that it spent two months leading a coalition of maintainers and security firms to address the bugs before public disclosure.

The scale of the disruption affected infrastructure and prominent web services globally. Wikipedia went down Wednesday morning, and the Wikimedia Foundation confirmed to Mashable that the outage was tied directly to the PHP vulnerabilities. Downdetector recorded issues at the same time for sites including Asana, GOG, Duolingo, Capital One, Bitbucket, Moodle, Quest Diagnostics, Final Fantasy XIV, War Thunder, and the Internet Archive. The vulnerabilities could affect 70% of the internet, allowing attackers to execute code and read memory containing sensitive information such as passwords.

Security firms and tech companies used frontier artificial intelligence models to unearth the flaws during recent security audits. Security firm Calif found one of the critical flaws using OpenAI’s GPT-5.6 Sol model as part of an OpenAI security program called Patch the Planet. Researchers at Wiz also discovered and reported both vulnerabilities using Google’s Gemini 4 Argon, Wiz Atlas, and OpenAI’s GPT-5.6 Sol. Wiz researchers stated in a blog post that a single researcher completed work in a week that previously would have taken a world-class team months.

The Fourteen Year History of the PHP2SHELL Flaw

According to details shared on X by Greg Brockman and reported by Mashable, the core bug had existed in PHP for 14 years. A smaller version of the vulnerability was found in 2024 but was not treated as serious at the time. In August, Calif discovered a version of the bug, and OpenAI’s internal team independently found a more complete version in September. The coalition originally planned to disclose the flaws on Oct. 22, but details leaked early, prompting PHP maintainers to release patches a day ahead of schedule.

Government Intervention and Global Response

Governments around the world mobilized emergency responses as the fixes went live. U.S. government agencies worked with OpenAI on its response, which involved the Cybersecurity and Infrastructure Security Agency, the Office of the National Cyber Director, and the White House’s Super Intelligence Force. The U.S. National Security Agency urged organizations using PHP to update their systems immediately or take them offline. In Japan, the prime minister issued an unusual public appeal on X, urging website operators to check their systems and install the patch released at 11 p.m. Japan time on Oct. 21.

Future Automated Defense Proposals from OpenAI

OpenAI stated that PHP will not be the last case of its kind and expects AI-driven vulnerability discoveries to rise sharply. Brockman noted that current bug-handling methods are slow and manual and will not scale as AI advances. To address this, OpenAI is asking the security community for feedback on a new disclosure coordination process. One proposal involves authorized defenders and governments constantly scanning the internet for vulnerable systems, warning site owners, and in extreme cases patching systems automatically. To support defenders against faster AI-powered attacks, OpenAI committed $1 billion toward subsidized access to its security tools for essential service organizations, and released a tool to help operators check for vulnerable PHP installations.

More on this story: OpenAI Models Breach Crypto Wallets and Disable Security in Tests

More on AI Models: Anthropic Warns AI Models Could Soon Self-Replicate Across Smart Devices · Telecom Operators Shift to Open AI Models, NVIDIA Report Finds