Hackers Deploy 44 AI Agents to Automate Attacks on Japanese Targets

A cybercrime operation deployed an automated attack framework coordinated by 44 AI agents to execute intrusions against Japanese entities, according to researchers who examined three exposed attacker infrastructures. The investigation follows a formal warning issued on October 9 by Japan's National Cybersecurity Office regarding unauthorized access and data theft across commercial networks.

Researchers at Trend Micro examined three attack environments to analyze how threat actors integrate artificial intelligence across the intrusion lifecycle. The analysts reconstructed the campaigns from publicly accessible directories left on remote servers, discovering attack scripts, target directories, command logs and operational reports.

Multi-Agent Architecture and CyberStrikeAI Abuse

The clearest intrusion case involved the abuse of CyberStrikeAI, an open-source penetration-testing platform adapted for malicious automation. Discovered configuration records such as credential-search agent eight of 44 indicated that the software divided work among 44 discrete algorithmic agents.

These software agents handled modular intrusion responsibilities, ranging from reconnaissance and vulnerability identification to exploitation, command and control, persistence, and data theft. The automated platform operated across multiple targets simultaneously while limiting network traffic against individual victims to evade detection. Furthermore, the platform generated Japanese-language forms and fake login pages to facilitate credential theft.

Target Validation and Network Reconnaissance Scope

Reconnaissance telemetry demonstrated broad operational reach across domestic networks, examining at least 6,000 Japanese IP addresses while probing older web platforms alongside computing clusters. Forensic analysis verified validation activity against at least 19 target organizations, including at least 16 institutions located in Japan.

In attacks targeting Yamaha RTX1210 routing hardware, the operators avoided zero-day exploits, relying instead on credentials recovered from configuration files rather than exploiting a product vulnerability. However, researchers noted that deep compromise affected only a small proportion of victims, with many server reports explicitly recording no session, no web shell, and no exfiltration.

A 24-day campaign in August and September 2026 targeted a geographic information system server using default credentials, according to The Hacker News.

Human Oversight Models and Linguistic Artifacts

Trend Micro researchers identified three distinct operational methodologies across the analyzed infrastructures, showing three different approaches ranging from limited AI assistance to systems using multiple automated agents. While Case A utilized AI to drive the full workflow, Case B was human-led with AI assistance in planning, and Case C showed AI use in tool development.

Linguistic artifacts showed consistent patterns across the command environments, where Simplified Chinese served as the working language in all three infrastructures. Security analysts concluded that algorithmic tooling primarily accelerated intrusion tempo, whereas initial breach mechanics continued to rely on well-known administrative oversights and unpatched servers.