Security researchers have uncovered severe vulnerabilities in PHP, the open-source scripting language that powers roughly 70% of the world's websites, with assistance from an OpenAI model.

The flaws, collectively dubbed PHP2Shell, have been described as potentially among the most severe discovered in PHP in more than a decade, and some experts have drawn comparisons to Heartbleed, the notorious OpenSSL vulnerability disclosed in 2014 that let attackers read chunks of server memory.

What the researchers found

The vulnerabilities surfaced during an audit conducted in August 2026 by researchers at Calif, a security firm. The team used GPT-5.6 Sol, one of OpenAI's models, to comb through PHP's source code. Ryan Austin of Calif is credited with discovering the vulnerabilities.

The bugs enable remote code execution and control over server memory. In practical terms, an attacker located anywhere on the internet could force a server to run instructions it was never meant to execute.

Part of what makes the finding significant is where the flaws sit: in PHP itself, rather than in any single application or plugin built on top of it.

According to the research findings, the flaws could expose sensitive data belonging to site visitors. They could also serve as a foothold for further cyber attacks against other systems.

How the disclosure unfolded

Calif CEO Thai Duong alerted Volker Dusch, who works on PHP, on August 11, initiating a quiet, coordinated effort to fix the problem before attackers could learn about it. That private-first approach is the standard security practice known as coordinated vulnerability disclosure, in which flaws are patched behind the scenes before details become public.

The response effort brought together PHP maintainers, outside researchers, security providers, and US agencies including CISA (the Cybersecurity and Infrastructure Security Agency) and ONCD (the Office of the National Cyber Director).

OpenAI announced its "Patch the Planet" initiative on October 21, 2026. The program is designed to mobilize defenders around the PHP2Shell vulnerabilities and to support the coordinated patching effort. A partial leak of information about the flaws forced the team to expedite the patch release to October 22.

As part of the initiative, OpenAI is contributing $1 billion in subsidized model access. The goal is to put AI tools in the hands of frontline defenders working on vulnerabilities like these.

Why AI-assisted bug hunting changes the picture

Experts quoted in the research drew the parallel to Heartbleed. They also stressed that AI-enhanced vulnerability discovery cuts both ways: the same capability that helps defenders spot holes could help attackers find and exploit them.

That tension sits at the center of OpenAI's $1 billion pledge. If models can surface critical flaws faster, the race becomes about who uses them first. Subsidizing access for defenders is an attempt to tilt that race toward the people doing the patching.

What this means

For website operators, the immediate priority is simple: update. A patch exists, and the partial leak means the clock started ticking before anyone wanted it to. Because PHP underpins such a large share of the web, that update applies far beyond any single site or hosting provider.

For the open-source world, PHP2Shell is another reminder of a structural problem: critical internet infrastructure often rests on projects maintained by small teams, a dependency that becomes a systemic risk when flaws of this severity emerge. The heavy involvement of US agencies in this disclosure shows how seriously governments now treat that dependency.

What to watch next: how quickly sites apply the October 22 patch, whether any exploitation of PHP2Shell is reported in the wild, and how OpenAI distributes its $1 billion in subsidized access.