AI Found a Landmine in the Web's Foundation — And It Affects Millions

The bomb that hasn’t exploded — yet

There is a particular kind of panic that comes from knowing a catastrophe is possible but not yet real. Construction sites in war-torn regions deal with this constantly. Unexploded ordnance sits in the soil, invisible until disrupted. The protocols are clear: stop work, alert authorities, wait for experts.

That is essentially the situation facing the global internet right now.

Less than two months after 100 technology companies signed an open letter warning about the need for cyber resilience in the age of artificial intelligence, the warning proved prescient in the most concrete way possible. An AI-assisted vulnerability scan discovered critical flaws in PHP — the programming language that powers roughly 70% of websites worldwide — and the fallout has been immediate, messy, and still unfolding.

How the discovery happened

The vulnerabilities were deeply embedded in the PHP codebase, sitting quietly for years. They are not trivial bugs that pop up from a casual code review. A cybersecurity researcher told the BBC he spent an entire night attempting to exploit one as a proof of concept and came away unsuccessful.

That is precisely the pattern researchers have observed before. Some of the most dangerous flaws in widely used software are stubbornly resistant to human discovery. They require a certain kind of persistence, intuition, and pattern recognition that even experienced engineers can miss. What they do not require is a team that can run millions of variations simultaneously, cross-reference patterns across decades of code, and flag anomalies that no single person would notice.

That is what AI brings to the table. Not a replacement for human judgment — but a multiplier for human effort at scale.

The PHP Foundation, which oversees the language, moved quickly once the discovery was confirmed. A patch was issued for the latest version of PHP. But the fix is not universal, and that is where the second-order consequences begin to pile up.

The leak that changed everything

Responsible vulnerability disclosure follows a well-established protocol. Researchers find a flaw, private notification goes to the software maintainers, patches are developed, and then — and only then — is the public alerted. The goal is to give defenders a head start over attackers. Ideally, a comfortable one.

This time, the protocol broke down. Details leaked before the patching window had closed. That means the clock started ticking with the entire world watching, and every adversary on the internet reading the same leaked information.

“There’s a real dilemma,” Professor Alan Woodward of the University of Surrey told the BBC. “How do you alert people they need to be ready, and that a patch is urgently required, whilst not alerting attackers?” His assessment was blunt: you tell everyone at the same time and hope the defenders are awake enough to act faster than the attackers.

From what is known, that race is still underway. Cloudflare and Palo Alto Networks reported millions of exploitation attempts in the hours and days following the leak. Whether any have succeeded remains unclear. Large website operators may not yet realize they have been targeted. Small sites running outdated PHP may be completely unaware.

Why this is different from a normal bug fix

A typical critical vulnerability — even a serious one — tends to affect a specific application or a narrow set of configurations. This one cuts to the foundation. PHP has been the backbone of the internet’s content layer for over two decades. WordPress, Drupal, Joomla, and countless custom applications all depend on it. The language is not going away. Its installed base is enormous and deeply entrenched.

The patch only applies to the latest version of PHP. Older installations cannot simply apply a fix. They need to upgrade first. For a well-maintained site running modern PHP, that is a straightforward process. For the vast number of sites still running legacy versions — some of them perhaps on servers that haven’t been touched in years — it is a significant engineering project.

This is where the human element becomes the bottleneck. Every website owner who is not technically sophisticated now faces a choice: delay the upgrade and risk exposure, or invest time and money into a system refresh they may not have budgeted for. Small businesses, nonprofits, and individual bloggers are the most exposed. They tend to be the least prepared.

The AI question

Gina Neff, Professor of Responsible AI at Queen Mary University, called the episode a success story in one respect. “We’ve seen how a team of people supported by a large AI company have used AI to find the problem, work for months together to solve it, and get everybody ready to be safe and secure,” she said.

That framing is not wrong. AI detected a flaw that humans missed. Defenders had a chance to respond before the attack materialized. No confirmed breaches have been reported. By one measure, the system worked.

But Neff also noted the mirror side of that coin: the same AI tools are available to adversaries. Those tools are likely already running, generating exploit paths, and testing variations against live targets. The race between defense and offense is no longer about who finds the bug first. It is about who can weaponize it fastest.

That dynamic — AI-assisted discovery speeding up both protection and exploitation — is the defining cyber-security challenge of the decade. This PHP episode is not an isolated incident. It is a prototype for what will happen again, and again, as these capabilities mature.

What happens next

The immediate pressure is on website operators to upgrade. The longer the window between leak and full patch adoption, the wider the exposure. Cybersecurity authorities around the world have issued urgent guidance. But guidance is not compliance.

On the infrastructure side, the PHP Foundation and the broader open-source community will face intense scrutiny. The question of how to prevent disclosure leaks in the future — especially as AI tools make vulnerability discovery faster and more accessible — is likely to dominate upcoming security conferences and policy discussions.

For the average website owner, the takeaway is stark: legacy software is no longer just a technical debt problem. It is a security liability with measurable, real-world risk. The unexploded bomb analogy holds because the danger is not abstract. It is sitting in the ground, waiting for someone to trip it.

Whether that person is a researcher trying to prove a point, a criminal actor trying to prove a business model, or an AI algorithm running experiments at scale — the trigger mechanism is becoming indistinguishable.