French government takes sites offline after AI discovers critical PHP vulnerabilities

black computer keyboard — Alexander Sinn / Unsplash

Since 19 October, cybersecurity alerts have followed one another, leading some to speak of a "black October" on social networks. Behind this noise, one subject keeps returning: artificial intelligence. It finds software flaws faster than before. It also helps pirates attack in mass using AI agents, programs that carry out all steps of an attack alone, from finding a flaw to intrusion.

On 21 October in the afternoon, internet users looking for their prefecture website thus landed on a maintenance page, just like those going to impots.gouv or Météo-France. The CERT-FR, alert centre of the ANSSI (the agency in charge of the State's cybersecurity), had just published an alert on PHP2Shell.

This nickname designates two flaws in PHP, the language that creates nearly 70 percent of web pages (WordPress, many online shops, many public sites). They allow taking control of a server from a distance, without a password. The day before, the European Union had already sounded the alarm.

At the request of the ANSSI, the Interior Ministry carried out preventive maintenance on all its internet sites exposed to a critical security vulnerability.

None of the cut sites had been hacked. Nevertheless, the State preferred to disconnect without waiting: more than 600 000 servers use PHP on the French Internet, according to the ANSSI, cited by ActuIA. These flaws slept in the code for more than 10 years, and it is the AI that allowed them to be found. Wiz, one of the companies behind the discovery, classifies them "among the most serious vulnerabilities ever found in PHP, and probably in the history of software", on its blog.

In the 10 hours before the public announcement, we saw thousands of exploitation attempts on our network. We passed to tens of thousands in the 45 minutes following the announcement, to hundreds of thousands in 2 hours, and to millions now.

These figures from the Cloudflare chief, a company that protects a large part of the world's sites, were picked up by Reuters. However, an attempt is still far from an intrusion: as of 22 October, the US cybersecurity agency CISA had not classified PHP2Shell among the flaws actually exploited.

That same week, another Web pillar wobbled. eMudhra, an Indian company that delivers HTTPS certificates (the little padlock that guarantees you are on the real site), protected eight of them with secret keys reduced to numbers below 256. One was simply worth 4. Anyone could have made a fake bank site with a proper padlock, before browsers blocked these certificates. On the same day as PHP2Shell, GitLab, a platform where companies store their software code, corrected a flaw rated 10 on 10, present since 2014.

Face to this influx, major AI companies bet on the same weapon for defence. Anthropic has thus widened access to Claude Mythos, a model so powerful that the company refuses to make it public.

Cyberattacks and AI agents: when artificial intelligence automates hacking

Finding flaws is only half the equation; their exploitation also automates. On 22 September, the security company Gambit reconstructed a pirate campaign who confided almost all the work to three free and reusable AI tools. Between 10 and 15 September, he launched 105 attacks and compromised at least 27 companies, for an average of a little over 25 dollars per target, according to the Gambit report.

More than 600 000 bank cards were stolen, and in one case, the agent's own cleaning routine destroyed the victim's data.

The cards came from two companies. Total estimated cost of the operation: between 12 000 and 18 000 dollars, the price of a used car for a haul of several hundred thousand cards.

On 19 October, NorForge opened the curtains of Cairn, one of these tools. Its researcher Snorre Fagerland spotted more than 20 dashboards, those from where pirates pilot their agents, left unprotected. He thus observed attacks in real time. The most active group, sinophone, led more than 1100 attack projects, mostly against online shops, for about 5 dollars per attempt and a success rate close to 10 percent.

Reuters recorded more than 100 victim companies, including John Deere, SharkNinja and Cummins. At the end of September, a tool of the same type was used against South Korean banks. CrowdStrike found on open servers the traces of ARTEX, a Chinese AI agent tool designed originally for intrusion tests (attack simulations commissioned by companies to find their own flaws). Data was stolen, including at Shinhan Bank.