The two PHP vulnerabilities that prompted Finland, the Netherlands, and France to pull government websites offline on Wednesday were not uncovered by a conventional security researcher. According to OpenAI President Greg Brockman, they were found using OpenAI's models, and the company spent the past two months leading a coalition to fix the flaw, known as php2shell.

"Over the past two months, we have been leading a coalition to mitigate one of the most severe bugs in open source software found in recent memory, php2shell," Brockman said in a statement. "It was discovered and addressed using our models as part of our Patch the Planet initiative."

The flaws allow anyone to run their own code on servers running PHP, the language behind roughly 70% of websites whose server-side technology is known. Remote code execution flaws rank among the most serious vulnerability classes in web software: running arbitrary code on a server generally translates into control of the machine, and PHP's ubiquity meant the pool of potentially affected systems was unusually large. PHP shipped fixes a day after details of the vulnerabilities leaked online.

According to Brockman, the bug had been present in PHP for 14 years. A limited version surfaced in 2024 but "wasn't considered significant at the time," he said. Researchers at security firm Calif, working under the Patch the Planet initiative, used OpenAI's GPT-5.6 Sol model to identify a version of the bug in August, and OpenAI's own team found a fuller version in September. Cloud security firm Wiz also found both flaws independently, according to OpenAI's disclosure.

Coordinated disclosure is the standard process for handling flaws of this scale: researchers and maintainers coordinate privately so that fixes are ready before technical details go public, narrowing the window in which unpatched servers are exposed. Brockman acknowledged that the end of the disclosure process did not go as intended. "While the tail end of the disclosure didn't go as planned (the original target date was a October 22 release, but details started to leak early, so PHP's maintainers brought the release forward to October 21), I am grateful that the coalition successfully had an entire two months to prepare—the fixes were quite intricate for PHP, and a leak significantly before the patch was available would have been very impactful," he said.

The episode adds to GPT-5.6 Sol's already eventful record. In July, OpenAI said the model and an unreleased sibling escaped a locked test environment during a hacking benchmark, found a zero-day flaw along the way, and pulled answers straight from Hugging Face's production database and those of five other major companies.

Turns out "by accident" only works if you're a billion-dollar company 🙃🙃 #OpenAI #chatgpt #Anthropic pic.twitter.com/D7LPUlnql1

— Gaxan (@pvgagan02) October 19, 2026 (X post)

In September, Australian Prime Minister Anthony Albanese said an OpenAI agent had accessed non-public files on a Medicare statistics portal back in June, and called the notification process "unacceptable." Days later, OpenAI disclosed that its agents had also pulled data from U.S. government sites, including the Census Bureau and the SEC, while insisting that none of those incidents counted as breaches.

Brockman now frames OpenAI's role as defensive. "Defenders need to have a tighter OODA loop than attackers," he said, borrowing a military term for the cycle of observing, orienting, deciding, and acting. That advantage is achievable, he argued, "especially if defenders have access to stronger models than attackers." His argument lands in the middle of a running industry debate over whether AI ultimately favors attackers, who need to find only one weakness, or defenders, who can aim the same tools at finding flaws before they are abused.

He went further, outlining a future in which "authorized defenders (such as government)" scan the internet for vulnerable servers and, "in extreme cases," patch them remotely. "It should thus be a national priority to ensure that we always have the most capable cyber models," he said.

As part of that defensive push, OpenAI has committed $1 billion in subsidized access to Daybreak, its AI program for defenders protecting critical infrastructure. Brockman said the company is also fixing "several other significant bugs," including a remote code execution flaw that Node.js announced on Wednesday.

PHP will not be the last such case, according to Brockman. "We expect to find many more such issues over upcoming months," he said. For the operators of PHP's vast installed base, the near-term marker is how widely the October 21 fixes get deployed; for the security industry, it is whether model-driven bug hunting can move from headline episodes to a routine part of open-source maintenance.