Skip to content
NewsIncident

OpenAI and Google's Wiz used AI agents to find a 22-year-old PHP flaw hitting millions of sites

· by Pondero Newsdesk

The short version

AI agents from OpenAI's Patch the Planet initiative and Google-owned Wiz found critical PHP vulnerabilities that had sat undetected since PHP 5 shipped in 2004, prompting an NSA warning to take unpatched sites offline.

OpenAI and Google's Wiz used AI agents to find a 22-year-old PHP flaw hitting millions of sites

Two critical PHP vulnerabilities sat undetected in every version of the language since PHP 5 shipped in 2004, until AI agents from OpenAI and Google-owned Wiz surfaced them this month, triggering emergency patches from the PHP Foundation and a rare US National Security Agency warning telling operators to pull unpatched sites offline if they cannot fix them right away.

What happened

PHP runs roughly 70% of the world's websites, including sites built on WordPress, and by some counts powers more than 33 million sites total, per TechCrunch's report. A security research firm called Calif found one of the two flaws using OpenAI's GPT-5.6 Sol model as part of OpenAI's "Patch the Planet" initiative. Separately, Wiz found and reported both vulnerabilities using its own Wiz Atlas security agent paired with Gemini 4 Argon and GPT-5.6 Sol, according to OpenAI's disclosure post. The bugs, which Wiz branded PHP2Shell, let an attacker reach a server's file system or an application's database, per Wiz's technical writeup. The PHP Foundation shipped fixes and recommends operators update to PHP 8.5.12; PHP 8.1 and earlier releases no longer receive security support, so some operators face a forced version upgrade rather than a quick patch. OpenAI said the coordinated response took roughly two months to prepare, per its disclosure post, running fixes through adversarial testing with frontier models to look for variants of the same bug before anything went public.

The exploit barrier that AI just removed

What kept these flaws dormant for two decades was not obscurity but difficulty: adapting a working exploit to the memory layout, configuration, and version quirks of each individual PHP deployment used to require enough specialist effort that attackers rarely bothered at internet scale. Wiz said that calculation has changed. "With AI this is no longer a barrier. LLMs can help adapt an exploit to different PHP versions, configurations and memory layouts," the company wrote, adding that AI tools can analyze failed attempts and iterate on adjustments automatically, which "significantly lowers the barrier to scaling attacks across different deployments," per Wiz's blog post. OpenAI said its researchers and Wiz's each found the bugs before any attacker did, a gap it calls a "Defender's Window." That window only stays open until the first working exploit appears in the wild, and the public disclosure itself hands attackers the same starting point defenders had. That is the practical reason the NSA told site operators to take a system offline entirely rather than leave it exposed while a patch is pending.

Context and reactions

Wiz credited Automattic, Akamai, AWS, Cloudflare, Fastly, Imperva, Microsoft, and Vercel with helping test the fix before the public disclosure, per OpenAI's post. WordPress, whose CMS runs on PHP, said the flaw is not specific to WordPress and told site owners to patch their servers and coordinate with hosting providers rather than wait for a WordPress-side fix, per TechCrunch's report. OpenAI said it also worked with multiple parts of the US government on the response, including the White House's Super Intelligence Force, the Office of the National Cyber Director, and the Cybersecurity and Infrastructure Security Agency, along with unnamed international partners, per its disclosure post. OpenAI used the disclosure to announce $1 billion in subsidized access to its Daybreak security product for organizations with limited security budgets that protect essential services, through a program it calls Daybreak for Frontline Defenders. It also published a free PHP version checker on GitHub, packaged as a Codex skill, so operators can identify which of their own systems are exposed before deciding whether to patch or pull a site offline. Separately, OpenAI said it funds security work for open-source maintainers more broadly through free Codex subscriptions and direct sponsorship of projects such as the Rust Foundation, framing the PHP response as one instance of a wider maintainer-support effort rather than a one-off.

What to watch next

OpenAI said it expects AI-assisted vulnerability discoveries like this one to increase sharply in the coming months and is asking the security community to weigh in on a formal disclosure framework it proposed alongside the PHP fix, covering how much advance notice maintainers and operators need before a bug becomes public. Watch whether other decades-old, widely used languages and frameworks get the same AI-driven sweep, and whether PHP operators who ignored the NSA's offline guidance turn up in a follow-up breach report once the exploit code catches up to the disclosure.

Sources