PHP Armageddon: What AI's Bug Hunt Means for the Internet

The internet just got a reminder that its backbone is ancient.

Two critical vulnerabilities in PHP — the server-side programming language behind roughly 70% of websites, or more than 33 million properties — have been discovered and patched, thanks not to a team of human security researchers working in isolation, but to AI agents from OpenAI and Google’s Wiz. The U.S. National Security Agency responded with unusually blunt guidance: if you cannot patch immediately, take the site offline.

The real story here is not the bugs themselves. It is what their discovery means for how the internet will be secured — or unsecured — going forward. This moment marks a tectonic shift in the relationship between AI and cybersecurity, one that will ripple through every layer of the web infrastructure stack for years to come.

Decades of rust, found in days

PHP 5 was released in 2004. These two flaws lived inside every version since, tucked into code paths that billions of lines of software touch but few humans have reason to examine closely. They allow an attacker to infiltrate a server’s file system or a database that a web application relies on. That is about as bad as it gets.

The patches are already out. PHP 8.5.12 is the current recommended version. But the reality on the ground is messier: sites still running PHP 8.1 and older are on unsupported releases. There is no way to know precisely how many of those 33 million websites are still sitting on vulnerable versions. The estimate from the PHP Foundation alone suggests tens of millions may need urgent attention.

The NSA did not mince words about the risk level. Its recommendation to pull unpatched sites offline is severe for a language this embedded in the web stack. But the agency’s urgency also reveals something uncomfortable: despite decades of automated security scanning, these vulnerabilities hid in plain sight for nearly twenty years. Human auditors missed them. Static analysis tools missed them. It took an AI agent — a system designed to reason through code at scale — to find what generations of security researchers could not.

AI found the bugs. AI will also lower the barrier to exploiting them.

This is where the story gets complicated.

Calif, a security testing company, was working within OpenAI’s “Patch the Planet” initiative to find flaws in open-source software when it used GPT-5.6 Sol to uncover one of the vulnerabilities. Separately, Wiz deployed its Wiz Atlas security AI agent running Gemini 4 Argon and GPT-5.6 Sol to find multiple PHP flaws. Both teams followed responsible disclosure practices, working with the PHP Foundation on patches before going public.

OpenAI confirmed that the flaws were found by researchers before any known exploits appeared in the wild. That is the good news. The bad news is what Wiz wrote in its own blog about the discoveries: before AI, these flaws were too hard for hackers to exploit. “LLMs can help adapt an exploit to different PHP versions, configurations and memory layouts,” Wiz warned.

In other words, AI is simultaneously the discovery tool and the enabler. It found the vulnerabilities first, which is exactly how this should work. But it also lowers the skill floor for turning a disclosed flaw into a working attack. The race is no longer between human researchers and human hackers. It is between AI-assisted defenders and AI-assisted attackers — and defenders are starting a patching campaign against tens of millions of sites that were never designed to be patched quickly.

This dual-use dynamic will define the next era of cybersecurity. Every AI system that discovers vulnerabilities faster will also accelerate the creation of weaponized exploits. The timeline between discovery and exploitation is shrinking toward zero, and organizations that rely on legacy infrastructure will bear the brunt of that compression.

The supply chain nobody talks about

Most public discussion of software supply chain risk focuses on dependencies — a malicious package in npm, a compromised build pipeline, a poisoned library. PHP vulnerabilities are a different category of risk. They sit at the infrastructure layer, beneath the applications and frameworks built on top of them.

The list of companies Wiz thanked for their help — Automattic, Akamai, AWS, Cloudflare, Fastly, Imperva, Microsoft, Vercel — reads like a who’s who of the web infrastructure stack. WordPress specifically issued a statement saying the flaws are not WordPress-specific. Sites running WordPress must patch their servers and coordinate with their hosting providers. That is the crux: most WordPress sites do not run on servers their owners directly control.

This creates a dependency chain that stretches from the PHP Foundation to hosting companies to site operators to end users. When a bug like this surfaces, the people who can fix it are often the least visible in the chain. Small businesses, nonprofits, and individual bloggers running outdated PHP versions on shared hosting plans are the soft underbelly here.

The second-order effects of this dependency problem are already visible. Shared hosting providers are bracing for a surge of support tickets. Enterprise IT teams are discovering that their legacy intranet applications depend on PHP versions that have reached end of life. Municipal governments and public institutions, which often run on outdated content management systems, face the same dilemma as small businesses — they cannot simply update software that has been stable for years, even if that stability now carries unacceptable risk.

The economics of forgotten code

There is also an economic dimension to this crisis that extends far beyond security. An estimated 60 to 70 percent of websites still run on PHP, and a significant portion of those are on versions that no longer receive security updates. The cost of updating — not just technically but organizationally — is prohibitive for many operators. Migrating from an old PHP version often requires rewriting application code, testing against new behavior, and retraining staff. For a small restaurant with a thirty-page WordPress site running PHP 7.4, that migration is not a weekend project. It is a budget decision.

Hosting providers are caught in the middle. Many shared hosting plans still default to PHP 8.0 or 8.1, versions that have passed end-of-life dates. Upgrading defaults across millions of accounts is operationally complex and financially risky — a breaking change on one customer’s site can trigger a wave of churn. Some providers are moving aggressively. Others are waiting, calculating the probability of an actual attack against the probability of customer dissatisfaction with forced migrations.

The NSA’s directive to take vulnerable sites offline exacerbates the tension. For a small business, going offline means losing revenue. For a hospital or emergency service, it means risking patient safety. The policy is technically sound but socially costly, and that gap between technical correctness and practical feasibility will widen as AI-driven discovery makes vulnerability disclosure more frequent.

What happens next

The immediate aftermath will be noisy. Security teams will scramble. Hosting providers will field support tickets. Some sites will go offline as operators follow the NSA’s conservative advice. The patching rush will reveal how much of the internet is running on software that has not been updated in years.

But the longer-term implication is more significant. AI-driven vulnerability discovery is no longer theoretical. It happened with two major flaws in one of the internet’s most widely deployed languages, in a coordinated effort that would have been nearly impossible at this speed and scale without large language models doing the heavy lifting of code analysis.

This will become routine. Security teams at companies like Calif and Wiz now have a playbook. OpenAI has an initiative dedicated to it. The question is whether the pace of AI discovery will outstrip the pace of AI-enabled exploitation, or whether we are entering an era where the most dangerous bugs are the ones found first by machines — and then, almost immediately, weaponized by other machines.

The PHP armageddon averted today is a signal, not an ending. The internet’s oldest infrastructure just got a wake-up call, and the alarm was rung by AI. The machines that found these flaws will find more. The question is whether humanity can build the institutional and economic scaffolding fast enough to keep the sites running when they do.