OpenAI launches Patch the Planet after GPT-5.6 catches critical PHP bugs

Close to 70 percent of the web just narrowly survived an existential scare.

A pair of critical flaws uncovered in PHP, the scripting language quietly running Wikipedia, WordPress, and countless web properties, sent system administrators around the world into emergency mode. Dubbed PHP2Shell, the vulnerabilities affect virtually every release of the language over the past 20 years, allowing attackers to execute arbitrary code on unpatched machines.

Wiz described the findings, cataloged as CVE-2026-109618 and CVE-2026-109619, as among the most severe security vulnerabilities in software history.

A 20-year-old landmine surfaced by AI

The most fascinating detail of the whole ordeal is how these bugs were uncovered in the first place.

OpenAI revealed that a researcher using GPT-5.6 Sol flagged the critical flaw through its Patch the Planet project. Cybersecurity firm Wiz independently caught and reported both vulnerabilities around the same time.

Recognizing the catastrophic fallout if exposed early, OpenAI spent the past two months spearheading a coordinated coalition. They quietly linked upstream PHP maintainers with Cloudflare, CISA, and the White House Super Intelligence Force to build and deploy web application firewall rules before public disclosure.

Yet attackers were already breathing down their necks. Amazon CISO CJ Moses revealed that their threat intelligence team caught an active PHP2Shell exploit attempt in their honeytokens on October 17th, days before the public patch went live.

Once disclosed, the floodgates opened. Cloudflare CEO Matthew Prince stated that malicious attempts on their network jumped from thousands in the hours prior to disclosure, to hundreds of thousands within two hours, and into the millions by nightfall.

Scrambled patches and lost digital history

While corporate giants scrambled, saving the web largely fell onto the shoulders of a single human.

Ondřej Surý, an open-source maintainer, ended up almost single-handedly packaging the patches that rescued the broader ecosystem. In an interview with The Atlantic, Surý admitted that while he remained oddly calm, AI-accelerated vulnerability discovery is taking an immense toll on underfunded frontline maintainers.

Sites like DeviantArt, 4chan, and parts of the Internet Archive briefly flipped the kill switch to apply emergency updates. Some Dutch and Finnish government agencies preemptively pulled web portals offline. Elizabeth Barron, executive director of the PHP Foundation, stressed that while no confirmed remote code executions had yet succeeded, patching is urgent.

For some legacy systems, however, this was a death sentence. Two long-running Japanese mobile games, World Quest and Save My Demonic World, announced permanent shutdowns because their decades-old backends could not be patched, illustrating how deep legacy dependencies threaten digital preservation.

OpenAI has since released a PHP version checker skill for Codex and pledged $1 billion in subsidized Daybreak access to help resource-strapped organizations defend vital services.

Why it matters

This incident provides an unsettling preview of the AI era in software security.

For developers and IT teams, the immediate takeaway is straightforward: check your stacks now. Relying on legacy, unmaintained PHP code is no longer just technical debt; it is an active welcome mat for high-frequency automated exploits.

For the broader industry, the balance of power has shifted. Frontier models like GPT-5.6 Sol can comb through 20 years of legacy codebases and find needle-in-a-haystack exploits faster than any human research team. OpenAI bought the ecosystem a critical defender's window this time, but Amazon's honeypot findings show adversaries are operating with near-zero lag.

We think OpenAI's structured disclosure approach was executed admirably, and committing $1 billion to frontline defenders addresses an obvious resource gap. But it also exposes a fragile reality: the modern internet relies heavily on underpaid volunteers like Surý to clean up fallout that AI tools can trigger in seconds. Unless critical infrastructure shifts rapidly toward memory-safe architectures, defending open-source software is bound to get much harder.