Machines · Theme of the week

One week, two warnings: a PHP flaw and AI attackers

Precautionary shutdowns across half of Europe and an AI attack campaign costing $5,000 show how fast and cheap attacks are becoming.

WORTH Signals · 24 Oct 2026 · 6 min read · Deutsch

One week, two warnings: a PHP flaw and AI attackers

This week two critical flaws in PHP led authorities, hospitals and universities in Europe, and banks in Japan, to take websites and online services offline as a precaution; no confirmed breaches had been reported by Friday. At the same time it emerged that a single attacker, using AI agents for about $5,000, targeted more than 1,200 websites and broke into over 100 companies.

The week

17 October: Amazon's honeypot servers record a first attempt to exploit the PHP flaws, four days before the official update. How the attackers obtained the information is unclear.
19 October: Security firm NorForge and Reuters make an AI-driven attack campaign public. Victims include websites of Deere and SharkNinja.
21 October: The PHP project releases the updates. Cloudflare counts tens of thousands of attack attempts within 45 minutes and millions by the following day. The BSI calls for immediate action.
21 and 22 October: In Germany, sites of the Federal Police, the Robert Koch Institute and customs, several state portals, cities such as Munich and Cologne, the Charité hospital and many universities go offline as a precaution. In the Netherlands, the DigiD citizen portal is down for several hours. In Japan, Shizuoka Bank, The 77 Bank and Mizuho suspend parts of their online services.
23 October: Most German services are back online. Several security firms still report no confirmed takeovers; BSI vice-president Thomas Caspers speaks of observed but unsuccessful attacks.

What happened

The week shows two developments that belong together. First, a single flaw in a basic technology is enough to take public services offline in many countries at once. The shutdowns were precautionary, not the result of successful attacks. But they show how little room operators have when attackers react within minutes and were apparently active even before disclosure. Second, attacks are getting cheap. In the campaign described by NorForge, each successful intrusion cost about $50 in compute. The agents worked autonomously, just under a day per intrusion on average, according to Reuters apparently using DeepSeek and Claude Code. In South Korea, attackers in late September and early October simultaneously targeted banks, savings banks and capital-markets firms with AI-assisted tools. Germany's BSI warned that attackers could exploit new flaws very quickly with AI tools. That shrinks the window between a flaw becoming known and the first attack, the time defenders have to update.

Questions

Could a flaw like this cause a blackout?

On current evidence, no. What was affected were websites, portals and apps, the public shop window. Assessment: power grids are controlled by dedicated control systems that do not run on PHP web servers and are meant to be separated from the internet. This week's shutdowns were voluntary and temporary. The real risk lies elsewhere: a hijacked web server can become an entry point into an internal network if systems are poorly separated. Amazon therefore advises tracing suspicious activity back to at least 17 October, including on servers already updated.

What happens if banks' old core systems are attacked?

Many banks still run accounts and bookings on mainframes with programs written in COBOL, a language from the late 1950s. These systems are generally not directly reachable from the internet. The risk arises at the junctions: apps, interfaces, service providers. In South Korea, experts warn that functions bolted on over decades make known weaknesses hard to fix. AI cuts both ways: it helps attackers analyse old code, and banks such as KB Kookmin use it to translate COBOL into modern languages. No successful attack on a core banking system became known this week.

Were German banks affected?

No German bank with online banking switched off appears in the overviews reviewed. In Germany, those affected were mainly authorities, hospitals, universities, media and hosting providers. In Japan, by contrast, several banks suspended parts of their online services as a precaution.

Coming weeks

The coming weeks are about clean-up. The bottleneck is legacy systems: by Friday the hosting provider Mittwald had no fix for very old PHP versions. Because attackers were active before the update, operators must check whether servers were taken over unnoticed. Assessment: experience suggests such headlines are followed by phishing waves, for example fake security notices from banks or hosting providers.

One to three years

The cost per attack falls and the number of potential attackers rises. If an intrusion costs about $50, even small targets such as mid-sized firms, medical practices and municipalities become worthwhile. For companies, the question shifts from defence as such to speed: those who apply updates in days rather than hours are at a disadvantage. In the EU, NIS2 and, for financial firms, DORA already impose duties on reporting and resilience. Assessment: the week gives supervisors and insurers arguments to tighten audits and tie terms to response times.

Five to ten years

In the long run the question is whether AI helps attackers or defenders more. The same models that find flaws can check code before it ships and modernise old programs. The debate is open: Anthropic chief executive Dario Amodei considers restrictions on open models necessary; others argue that open models mainly strengthen defence. Assessment: for organisations with a lot of legacy software, including public authorities and banks, modernisation becomes a security issue, not just a cost issue. Those who understand their systems and can change them quickly have a structural advantage.

What to watch

Whether CISA or the BSI confirm first successful takeovers via the PHP flaws
Whether BaFin or the EBA question banks on AI-assisted attacks or issue requirements
Further reports of autonomous AI attack campaigns and how model providers respond