Trend Micro researchers reconstructed three attack environments left on publicly reachable servers and found AI sat at a different level of control in each.
It arrives during a rough autumn for Japanese organizations. On October 9, Japan’s National Cybersecurity Office warned businesses, and the ruling party’s cyber team head called cyberspace “a state of emergency.”
The team examined files operators abandoned: scripts, target lists, logs, command histories and reports. They could not see victim-side logs, so a result counted only where evidence supported it.
Three cases, three architectures
The most automated case used CyberStrikeAI, an open-source platform built on AI agents. It split work among 44 agents and carried reconnaissance, exploitation, command and control, persistence, theft and reporting as one repeatable pipeline. At least 19 organizations, 16 Japanese, were probed this way, each getting an identically formatted report.
The second was human-led, with a large language model accelerating planning and tool building. It ran 24 days across August and September and included an intrusion into a geographic information system server left on default credentials, yielding cadastral records and email.
The third confined AI to writing tooling. That actor targeted edge devices from Cisco, Citrix, F5 and Yamaha using published flaws, admin functions and credentials in config files, running at least six campaigns from January to October 2026.
Same doors, faster knocking
Trend found no evidence the three share an operator and could not identify the models used. Reach is not limited to Japan and South Korea: the first operator hit Taiwan and Hong Kong, the third kept WebLogic discovery lists naming UK, French, Italian, Qatari and Colombian targets.
“The types of doors and the weaknesses that leave them open have not changed,” the researchers wrote. What changed is the ability to try thousands at once and adapt to local language.
