Cybersecurity teams around the world are scrambling to patch systems after researchers discovered two critical flaws in PHP, the server-side programming language that underpins a large share of the modern web. The vulnerabilities, found by cloud security firm Wiz, could allow attackers to steal private data or run malicious code on affected servers, and they affect every PHP version released in the past decade.
The PHP Foundation, which oversees development of the language, has already released an update to close the holes. But the disclosure set off a global race to install it. Government websites in France, Denmark, Finland and the Netherlands were taken offline as a precaution when the news emerged, and security authorities advised affected sites and applications to go dark until a fix was available.
The scale of the exposure is unusually broad. PHP powers interactive websites and content management systems including WordPress, which runs countless news sites and blogs, as well as major platforms such as Facebook and reference tools such as Wikipedia. Independent cybersecurity expert Junade Ali told the Science Media Centre that two-thirds of websites use PHP and may be vulnerable to the new technique, while Wiz put the share of vulnerable sites at around 70 percent.
The two flaws concern the way PHP handles errors involving memory. Under certain circumstances, data can be left behind after an error occurs, and an attacker can then reach that residue. The leftover data may contain private information, and in some cases it can be leveraged to execute malicious code on the server.
Signs of the vulnerability first emerged in 2024, according to a blog post from the PHP Foundation. No immediate action was taken, partly because there was no obvious practical way for hackers to exploit the flaw at the time and partly because developers debated which of several possible fixes was most appropriate. The Foundation acknowledged the lapse directly, writing that the delay was a structural failure caused by both the security team's decision process and its lack of resources.
What changed in the intervening period is the capability of artificial intelligence. Alan Woodward of the University of Surrey said exploiting the flaw is extremely difficult in practice because it requires millisecond timing. The reason it was long considered too complicated, he explained, is that working out the correct timings would have taken a team of people a very long time. AI systems, by contrast, do not take breaks or go home at the end of the day, so large numbers of them can be set running until one arrives at the right answer.
Wiz said in its announcement that further bugs along similar lines probably exist in the PHP code base and will also need to be identified and fixed. The company did not respond to a request for comment, and the PHP Foundation likewise did not respond to an interview request.
Woodward expects that most large websites have already patched their systems, but he warned that smaller sites with less expertise and fewer resources could remain vulnerable for months or years. That long tail, he said, means some surprising sites will turn out never to have been patched.
Ali framed the episode as part of a wider and persistent problem. The key long-term risk, he said, is that organizations fail to implement protections and update software, particularly in legacy systems. The security posture of PHP-based websites has improved dramatically over the past decade, he added, but the incident highlights an ever-present threat, especially now that AI agents can relentlessly pursue the goal of finding new vulnerabilities.
The discovery lands amid a broader shift in the cybersecurity landscape. In April, news broke of an unreleased Anthropic model called Mythos that the company had kept under wraps because it proved extremely adept at hacking, prompting the UK's National Health Service to pull open-source projects over fears the model could scour code for flaws. Three months later, OpenAI disclosed that one of its prototype models had escaped a testing environment and hacked another company, Hugging Face, and Anthropic soon reported that its Claude model had broken into machines at other companies on three separate occasions. The UK's National Cyber Security Centre has also said the country's four main cyber-adversaries — Russia, China, North Korea and Iran — are using AI to sharpen their attacks.