Flaws identified using artificial intelligence (AI) in programming software widely used on the internet led to thousands of sites being taken offline on Wednesday as a precaution, an illustration of the new situation imposed by AI on cybersecurity.
The American intelligence agency NSA and the British National Cyber Security Center (NCSC) on Wednesday recommended that all sites using the PHP programming protocol disconnect before updating the software.
In Europe, many platforms have been suspended, notably those of all prefectures in France, at the request of the IT Systems Security Agency (Anssi).
In Finland, the government internet portal reported that “government websites (were) completely or partially unavailable.”
Public service and community platforms have also been taken offline in the United Kingdom, Germany, Denmark, the Netherlands and Belgium.
22 million platforms exposed because of PHP
Earlier on Wednesday, the PHP Foundation, which coordinates the development and evolution of PHP, published a note reporting two vulnerabilities.
PHP, short for Hypertext Preprocessor, is a programming language created in the early 1990s by Canadian-Danish programmer Rasmus Lerdorf. It allows you to establish a gateway between a website and a server housing databases.
Cybersecurity specialist Palo Alto Networks said on Wednesday that it had identified 22 million platforms exposed to these vulnerabilities.
“PHP powers about 70% of the global web,” described Sagi Tzadik, researcher at cybersecurity specialist Wiz, in a podcast published by the company on Wednesday.
Flaws discovered by AI
Also on Wednesday, OpenAI explained that the flaws were found by IT security company Calif using one of its most recent models, GPT-5.6 Sol.
“Given their severity and the wide distribution of PHP, we consider this to be one of the most impactful (computer) security issues in recent memory,” the start-up wrote in a message published on its site.
In addition to Calif, the cybersecurity company Wiz, a subsidiary of Google, also pointed out these shortcomings, without specifying which models it had used.
“This shows how AI can uncover breaches with consequences that affect the entire internet,” OpenAI added, while warning that these same capabilities “(could now) help (computer) hackers quickly develop (attacks).”
It is generally agreed that AI has taken a step forward this year in this regard, to the point that Anthropic has decided not to widely market its Mythos model, considering it too risky for cybersecurity.
AI for hackers
OpenAI’s main rival has only made it available to a handful of organizations to carry out software updates that fill the gaps spotted by Mythos.
The use of AI increases the potential for discovering cracks in software but also significantly reduces the reaction time of actors exposed to attacks. Thanks to artificial intelligence, hackers can target many more targets than before and within very limited time frames.
Wednesday’s incident follows another alert on Tuesday concerning an encryption protocol called TLS which guarantees the confidentiality and integrity of exchanges on the internet.
Several examples have been reported showing that hackers could impersonate the user and decrypt communications.
“Two categories of extremely serious vulnerabilities were revealed in 24 hours,” reacted, on X, Matthew Prince, boss of the IT service provider Cloudflare, which notably offers cybersecurity services.
“The first (TLS) affects everyone online,” he warned. “The second is almost everyone who uses a server.”
