European government sites closed one after another...AI found a 10-year-old security hole right away

A serious security flaw in PHP, a programming language, has shaken web systems around the world after artificial intelligence (AI) found it. Several European governments have temporarily shut down public websites as a safety measure. The US National Security Agency (NSA) advised taking systems offline if they could not be updated with a security fix right away. Experts say the incident shows that AI can find hidden security flaws that have existed for more than ten years, while the tools used to attack systems are also being developed much faster.

According to Reuters and Bloomberg on the 21st (local time), PHP developers released an urgent security update to fix two flaws known as "PHP2Shell." The flaws were given a critical severity score of 9.2 out of 10 under version 4.0 of the Common Vulnerability Scoring System (CVSS).

The flaws allow remote code execution (RCE). This means an attacker outside a server can make it run almost any code without logging in or proving their identity. The problem comes from defects in how the PHP engine handles running out of memory and errors. According to W3Techs, a company that studies web technology, about 75% of websites whose programming language is known use PHP. Major web platforms, including WordPress, MediaWiki, and Drupal, also use PHP.

AI tools were used to find the flaws. OpenAI said that Califf, a security company, found a critical flaw using OpenAI's "GPT-5.6 Sol" model. Wiz, a security company linked to Google, also found and reported the two flaws independently.

Records from the PHP Foundation show that Califf began checking PHP code on July 30 and reported a flaw that allowed remote code execution on August 11. OpenAI reported the same underlying problem separately on that day. Wiz said it used Google's "Jeminai 4 Argon" and GPT-5.6 Sol together. The company said one researcher developed attack code in about a week, a task that experts might normally need several months to complete. OpenAI called the incident "one of the most far-reaching security problems in recent years" and warned that AI's ability to find flaws could also help people create attack tools more quickly.

The public release process made the situation worse. The PHP development team had originally planned to release the security fix on the 22nd. But after the vulnerability ID number and planned release time were leaked on an online message board on the 20th, the team moved the release to 2 p.m. on the 21st (11 p.m. Coordinated Universal Time).

That same night, Califf, which had found the flaw, accidentally uploaded a video showing a proof of concept (PoC), demonstrating how the attack worked against a real target, and then deleted it. The PHP Foundation said copies of the video had already spread, helping people develop public attack code the following day. Then, at 7:41 a.m. on the 21st (Coordinated Universal Time), security company Chainguard released an updated PHP image more than six hours before the official release. Like other distributors, Chainguard had received the patch in advance. When a security fix is released, attackers can compare the code before and after the change to work out where the weakness was. The PHP Foundation criticized Chainguard for releasing it early, saying, "It ignored the work that several parties have been doing for weeks for joint disclosure." Dan Lorenc, Chainguard's CEO, said he chose to release it early after several security companies reported that real attacks had begun. The PHP Foundation, however, said it had not been able to confirm that working public attack code existed at that time.

Problems within the PHP Foundation itself also came to light. The foundation said it had noticed an early sign of one of the two flaws while responding to an outside security audit in 2024. However, work stopped for months because people could not agree on how to fix it. The foundation called this a "structural failure." It has since changed the way it makes security decisions so that a majority vote decides them, and it has hired engineers.

The effects quickly spread to public services. Reuters reported that the Netherlands' National Cyber Security Centre took its own website offline. The Port of Rotterdam, government websites, five major political parties, and some hospital websites also stopped operating. Government IT service agencies in Finland and local government IT companies in Denmark suspended services as well. In Denmark, national school communication platforms were also affected. France's Interior Ministry temporarily blocked access to regional government offices and local administration websites. Belgium and Canada also took some websites offline. Organizations that run Wikipedia shut down some systems, and Internet Archive temporarily went offline too. The NSA said on X, "If an immediate update is not possible, we recommend taking the system offline until it is fully acted upon."

The danger remains even after the patch has been released. According to internet analysis company Sensis, about 1.5 million of the roughly 2.6 million servers confirmed to be running PHP use versions that do not have a patched release available. PHP versions 8.1 and older no longer receive official support, so there is no official update for them. The development team released only the basic source-code changes needed to address the flaw in versions 7.4 through 8.1. Cloudflare said it is continuing to update its security rules after early attempts to exploit the flaws appeared.

South Korea's Internet & Security Agency (KISA) issued a recommendation to install the security update late on the 21st. As of the morning of the 22nd, there was no confirmed report of a government order requiring public institutions to inspect their systems.

OpenAI proposed a new process for working together to respond to security flaws found by AI on the same day. It also asked the industry to share its views on how long organizations should have to fix flaws before they are publicly disclosed, and on how emergency security updates should be distributed. [Silicon Valley correspondent Wonho-seop]