TrendAI analyzed three attacker server environments found as open directories, separate from the earlier ARTEX campaigns. Each held tools, target lists, logs and command histories, and each used Simplified Chinese.

• Case A, AI-driven (confirmed): the open-source agent pentest tool CyberStrikeAI, with records such as "credential-search agent 8/44" showing 44 agents splitting the work: recon, vulnerability checks, exploitation, C2, persistence, exfiltration and automatic reports. It also generated Japanese forms and fake login pages. In a short burst in October 2026 it ran validation against at least 19 organizations, at least 16 in Japan, plus targets in Taiwan and Hong Kong. Only a very small share showed deep compromise; the clearest one came in through an unauthenticated admin screen on an old web system.
• Case B, human-led with an LLM (confirmed): about 24 days in August and September, more than 6,000 Japanese IPs scanned, real damage at one or two companies, including theft of cadastral and commercial records, Windows credentials and email from a GIS server left on default credentials. It also targeted AI workflow platforms and GPU/inference services, apparently to get cheap inference.
• Case C, AI for tooling (high confidence): January to October 2026, exposed VPN devices including the Yamaha RTX1210 and at least 1,300 WebLogic candidates, mainly in Japan and South Korea.

Trend Micro's conclusion (its opinion): AI raised the scale, speed and language fit of these attacks, but the attackers were still "looking for unlocked doors": legacy systems, default credentials and exposed admin interfaces. It found no evidence the three share an operator, and has not published IOCs, giving them to law enforcement instead.