The vulnerabilities were uncovered by researchers at OpenAI, the cybersecurity company Wiz, and another security firm. The underlying weaknesses had been present in PHP for more than ten years, but turning them into a working attack demanded a chain of steps that sat, as Wiz put it, close to the edge of what human researchers could realistically pull off on their own.
That is where the machines came in. Researchers used frontier AI models to adapt known techniques, build working exploits, and read the results at speed. Wiz said a single researcher finished in about a week what a top-tier team would once have needed months to complete. The firm called PHP the backbone of the web and described the episode as an AI inflection point, a moment when the technology shifted from assisting defenders to reshaping how fast flaws can be understood and weaponised.
The AI cybersecurity threat here is not that a model wrote malicious code on its own. It is that AI compressed the time and skill needed to grasp a buried weakness and act on it.
What the vulnerability does
The critical PHP flaw involves how the language handles a computer’s memory. When a website request ends without warning, PHP can leave data sitting in memory, where it can later be reached, altered, or exposed. Cloudflare, the web infrastructure provider working with the PHP Foundation and the research teams, said the flaws could let attackers run code remotely, knock services offline, and pull customer data.
Patrick Fussell, global head of adversary simulation at IBM, offered a plain picture of the risk. He compared it to checking out of a hotel room while your key still opens the door, so you can walk back in later even though someone else’s belongings now fill the space.
The remote code execution flaw is the part that worries defenders most, because it hands an intruder the ability to act as though they belong on the server.
Who is exposed
Nigel Green, chief executive of financial advisory firm deVere Group, warned that the reach runs almost anywhere people spend money online. He pointed to booking platforms, marketplaces, subscription services, ticketing sites, delivery apps and utility portals, any of which might store card details on systems built with the same kind of code.
The scale is the point. With PHP sitting behind so much of the web, a single pair of flaws touches a share of the internet that few other bugs could match.
The response from operators and leaders
The PHP Foundation, the body behind the language, released a website security update and urged operators to apply it without delay. The bugs were found in late August and kept quiet until Wednesday, once a fix was ready. Cloudflare’s chief executive said on X that the company’s network had already logged millions of attempts to exploit the flaws within days.
Political leaders moved quickly. Dutch Prime Minister Rob Jetten said his cabinet was watching the situation closely. Japan’s Prime Minister Sanae Takaichi went further, telling companies to weigh suspending services where needed and asking operators to check whether their public sites run PHP and to prepare the patch at once.
Several large organizations followed that advice. In Japan, airline ANA, railway operator JR West, and the country’s largest bank, MUFG, reported some pages and services briefly suspended. Across Europe, government websites in France, Denmark, Finland and the Netherlands, including the Dutch cyber security agency, were pulled offline while fixes went in.
Services taken offline as a precaution
The PHP vulnerability affecting websites reached well-known institutions. The Wikimedia Foundation, which runs Wikipedia, said it was addressing the PHP flaws and had taken several systems down as a precaution. In the UK, government minister Elizabeth Lloyd told the House of Lords that nearly three dozen government services were unavailable and four partly unavailable while upgrades took place. The Journal, Ireland’s second most-popular news outlet, went offline on Thursday morning while it strengthened its defences.
For website operators, the practical task is narrow and urgent: confirm whether a site runs PHP, apply the patch, and watch for signs of compromise. The vulnerability patch closes the hole, but it cannot undo access that may already have occurred.
A comparison that frames the stakes
Wiz researchers placed the episode alongside Heartbleed, the 2014 bug that struck an encryption technology thought to protect around two-thirds of the web. Heartbleed let criminals reach data that was meant to stay sealed. Like that earlier event, the researchers argue, PHP2Shell could mark a shift in how cybersecurity works, with AI turning long-dormant flaws into practical threats.
That reframing matters for the people who set security rules. The question is no longer only which bugs exist, but how quickly anyone, defender or attacker, can understand and use them. Governance of AI tools, disclosure timelines, and the duty of operators to patch fast now sit at the centre of the story. The flaw is technical. The accountability is not.
