AI-Built Malware Has Left the Lab and Hit Japan
Three intrusions investigated by Trend Micro show criminals handing parts of the attack chain to language models. Microsoft and Palo Alto's Unit 42 say the shift is already changing how fast attacks move.
Key points
- Trend Micro analysed three attacks on Japanese organisations where AI either guided the hackers or ran parts of the operation, with one toolkit coordinating 44 separate AI agents.
- Unit 42 has found live malware samples written or steered by large language models during an attack, confirming a shift researchers warned about last year.
- Microsoft says AI is "compressing attack timelines," meaning the gap between a flaw being found and criminals exploiting it is closing.
- Threat Vectr's tracking of criminal leak sites recorded a steady rise through 2025 in gangs claiming AI-assisted intrusions, though most claims can't be independently verified.
- No evidence yet exists of fully autonomous malware making its own decisions on a victim's machine, according to Unit 42.
Researchers in Japan have pulled apart three recent cyber attacks and found something the industry has been bracing for. The hackers weren't working alone. They were working with, and in one case largely through, artificial intelligence.
The analysis comes from Trend Micro, which examined three exposed servers used to run the attacks. One toolkit, dubbed CyberStrikeAI, coordinates 44 separate AI agents: small programs that each handle one job, such as scanning a network, writing a phishing email, or picking the next target. That moves AI-driven attacks out of the proof-of-concept folder and into real victim networks.
What does "AI-assisted malware" actually mean?
It means criminals are using large language models, the technology behind tools like ChatGPT, to do work that used to need a skilled human. Unit 42, the research arm of Palo Alto Networks, splits the activity into three categories: writing the malware itself; using AI to make decisions mid-attack, for example picking which file to steal or which machine to hit next; and running an AI model directly on the victim's computer with no human in the loop. Unit 42 says it hasn't seen that third category in the wild yet.
The Japan samples sit in the first two. A human is still pulling the strings, but the grunt work has been handed off. We've been tracking Unit 42's findings closely: our 19 September report on credential-based intrusions found AI had already cut the time between a stolen password and a full network compromise.
How worried should ordinary people be?
Not panicked, but aware. Attacks still arrive the same way: a dodgy email, a stolen password, an unpatched server. What changes is the speed and the volume.
Microsoft's 2026 Digital Defense Report puts hard numbers on the scale. Its systems block 4.7 million new malicious files every day. Microsoft's own phrasing is that AI is "compressing attack timelines." In plain English: the gap between a flaw being discovered and criminals exploiting it is getting shorter, fast.
Who is doing this?
| Research source | What they found | Where |
|---|---|---|
| Trend Micro | Three live attacks using AI, one with 44 coordinated agents | Japan |
| Unit 42 (Palo Alto Networks) | Two malware samples written or steered by LLMs | Global OSINT hunt |
| Microsoft | AI shortening the time between flaw and exploit | Global telemetry |
Threat Vectr's tracking of criminal leak sites, the dark-web pages where ransomware gangs name their victims, logged a steady rise through 2025 in groups claiming AI helped them break in. Those claims are marketing as much as fact. The research from Trend Micro and Unit 42 shows the underlying capability is real. Japan's exposure here isn't new, either: our 8 October story on the country's data leak surge showed attackers were already moving fast through unpatched systems before AI gave them another accelerant.
Should you do anything differently?
The basics still work. Turn on multi-factor authentication, the second-step login code, on email and banking accounts. Keep phones and laptops updated. Treat unexpected messages asking for money or passwords with suspicion, because convincing is exactly what AI is now optimised for.
For businesses, the honest read is this: if your security plan still assumes attackers work at human speed, it's already out of date. The Japan cases aren't a warning about what's coming. They're a report on what's arrived.
