Skip to the stories
Windows & PowerShell

Frontier AI surfaces two severe PHP flaws behind a large share of websites

Two critical PHP flaws rated 9.2 on CVSS v4.0 could let attackers run code on web servers. Here's what admins should check now.

Frontier AI surfaces two severe PHP flaws behind a large share of websites

Wiz, the Israeli cybersecurity firm now part of Alphabet, has called these two PHP flaws among the most severe vulnerabilities found in the language's history, and possibly in software generally. That's a strong claim, and it's the reason sysadmins should read past the headline. Some websites were pulled offline on 21 October to shut down exposure while patches rolled out, and the discovery itself came from frontier AI models.

What was found and how bad it is

PHP is a server-side scripting language built for web development. It handles the backend logic and talks to databases before the finished page reaches a visitor's browser. Nearly 70% of websites globally run on it, according to the figures reported in this coverage, which makes a flaw in the core language a problem for far more than a handful of niche applications.

The two critical flaws are rated 9.2 out of 10 on CVSS v4.0, the industry severity scale. Cloudflare, which reported on them, said they could let attackers execute malicious code on vulnerable servers, expose customer data, or disrupt services. It also pointed out that WordPress and Magento are built on PHP, so the blast radius reaches well beyond the language itself. If you manage a fleet of CMS-driven sites, you're probably looking at a larger share of your estate than you'd expect.

Household names co-sign open letter on AI cybersecurity threat

Wiz put it in its own terms: "The underlying flaws had been present for decades, but demonstrating their exploitability required an exceptionally complex chain of steps…frontier AI models substantially reduced that barrier, turning a longstanding flaw into an exploit." That sentence matters more than the score. It says the bugs weren't new, and that what changed was the cost of turning them into a working attack.

Why AI changed the maths on this one

For most of the history of this kind of bug, a flaw like this sat in the code for years because proving it could be exploited took specialist time and patience. Chaining several steps together to get from a weakness to actual control of a server is slow, skilled work. Wiz's point is that frontier AI models shrank that effort enough to move a latent flaw into the category of practical exploit.

OpenAI says it won’t release new GPT-6.1 Astra over safety concerns

That's a shift worth taking seriously even if you never touch PHP directly. It means the window between a bug existing and a bug being weaponised may be getting shorter across the board. Teams that treat patch cycles as monthly rituals should expect that assumption to be tested more often.

The discovery trail is also worth knowing. The organisation first found the flaws in August, working with a group of researchers and security providers that OpenAI described as part of its Daybreak initiative. A security firm called Calif identified one critical flaw, and Wiz independently discovered and reported both. PHP's maintainers then issued a warning and led the upgrades. The sequence suggests a coordinated disclosure rather than a leak, which is the outcome most people would hope for, but it doesn't reduce the urgency for anyone still running unpatched builds.

Geoffrey Hinton among 415 scientists calling for joint effort for AI safety
PHP code on a computer monitor. Close up.

Image: © Minerva Studio/Stock.adobe.com

What an unpatched server would expose

ESET Ireland warned that, left unpatched, the vulnerabilities could allow an attacker to take control of a web server without needing a password. That's the scenario that should drive prioritisation. A password-free takeover of a web host isn't a data-leak risk alone. It's a foothold from which an attacker can alter content, serve malicious code to visitors, or move into adjacent systems.

OpenAI agents tamper with US government sites after Australian breach

The public-facing risk is subtler, and this is where the ESET comments get practical. George Foley, a cybersecurity specialist at ESET Ireland, said: "Most Irish small business owners have never heard of PHP, but there's a good chance their website runs on it." When an attacker quietly takes over a genuine site, they can push malware or fake login pages to visitors who trust that domain. Trusted-domain abuse is harder to spot than a random suspicious link, and it's why a site that looks completely normal may still be compromised.

Expect the social-engineering side to follow the technical story. Foley warned that emails telling businesses their site has been hacked and needs an urgent fix through a link are close to a certainty. The safe pattern is to go to your web provider or hosting platform directly, not through any link in an unsolicited message. That advice applies to customers as much as to the people who run the sites.

OpenAI agents breach Modal client system after Hugging Face hack

What IT teams should check this week

The source material doesn't include version numbers, patch identifiers or step-by-step remediation, so I won't guess at any of them. What it does make clear is the sequence to follow:

First, build an inventory of where PHP runs in your estate. That includes obvious web servers, but also CMS platforms such as WordPress and Magento, hosted sites managed by third parties, and internal tools you may have forgotten about. Many organisations discover PHP in places nobody documented.

Nvidia forms industry alliance after Hugging Face cyber incident

Second, confirm patch status with the actual vendor or maintainer for each platform. The ESET guidance is to contact your web developer or hosting provider to confirm whether sites have been updated. For managed hosting, that conversation matters more than your own patch dashboard, because the provider controls the runtime.

Third, treat unexplained changes as a possible indicator. Look for new admin accounts, altered files, unfamiliar scripts or redirects, and login pages that differ from what you deployed. If the flaws were exploited against a site you run, those are the places attackers typically leave evidence. Don't run ad hoc fixes on a live server without a tested rollback, especially when the change involves a web-facing runtime.

Fourth, brief the people who handle customer-facing communications. A compromised site often shows up first as customer complaints about strange pop-ups or downloads. Having a short script for staff, and a rule that password and payment requests are verified through known channels, will save time if the worst happens.

What to watch next

My read is that the bigger story here isn't the PHP flaws alone. It's the demonstration that AI models can lower the bar for exploiting code that has sat in production for years. Whether that holds up will depend on how quickly the ecosystem patches, how many legacy sites never get updated, and whether similar findings start appearing across other open-source languages and frameworks.

For now, the job is straightforward. Find every PHP runtime you're responsible for, confirm the patch state with the people who control it, and keep an eye on the email and phishing traffic that tends to follow a disclosure like this. The flaws are old, the exploit path is newer, and the gap between those two facts is where most of the risk sits.

Source: Suhasini Srinivasaragavan · www.siliconrepublic.com

Same flag