Offensive Artificial Intelligence — Automated Campaigns Using CyberStrikeAI and 44 AI Agents in Operations Against Japan

Publication date: October 27, 2026
Category: News

Introduction

Security researchers at TrendAI have disclosed critical details regarding multiple cyberattack campaigns targeting organizations in Japan and other Asian countries, where artificial intelligence (AI) serves as a core component in automating the attack lifecycle. Unlike previous incidents leveraging autonomous penetration testing frameworks like ARTEX, the newly analyzed operations utilize dedicated workspaces and AI agent platforms—most notably the open-source tool CyberStrikeAI—to manage comprehensive workflows ranging from initial reconnaissance to data exfiltration and structured report generation. This finding highlights how adversaries are integrating language models and cooperative agents to scale malicious operations with unprecedented speed and autonomy.

What Is Agent-Based Offensive Automation? (General Analysis)

Agent-based offensive automation represents an evolution in malicious cyber operations, combining large language model (LLM) capabilities with autonomous agent frameworks to drastically reduce the operational cost, time, and technical expertise required to breach systems. Rather than relying on human operators to execute manual commands, these platforms integrate dozens of specialized agents that collaborate to fulfill tactical objectives.

Because this threat does not correspond to a specific software vulnerability with a dedicated identifier (CVE), there is no official CVE or CVSS score assigned; however, from a risk management perspective, the impact of these platforms aligns with CWE-284 (Improper Access Control) and CWE-200 (Exposure of Sensitive Information to an Unauthorized Actor) due to the systematic exploitation of default configurations and exposed services. The adoption of these systems compresses the time window between asset exposure and effective exploitation to well below 24 hours.

How Does It Work? (Technical Analysis)

Analysis of the exposed environments discovered by TrendAI and other research teams outlines a highly structured and modular operational ecosystem:

  • Multi-Agent Delegation: In the specific case of CyberStrikeAI, the framework delegates critical tasks across a pool of 44 specialized agents. These agents independently handle aspects such as vulnerability discovery, exploit validation, persistence, and the generation of localized phishing and fake authentication pages in the target’s native language (Japanese).
  • Automated Attack Lifecycle and Reporting: The workflow operates repetitively without human fatigue. Each evaluated organization undergoes structured processing that culminates in an identical standardized report and an evidence index automatically generated by the platform.
  • Additional Operational Variants:
    • Human-led LLM-assisted cases: Human operators utilize language models to plan attacks against AI workflow platforms and inference services, successfully stealing Windows credentials, email data, and cadastral records after compromising a GIS server configured with default credentials.
    • Specialized Tooling: Other threat actors employ models and frameworks such as Hermes Agent (powered by DeepSeek V4), Claude-based multi-agent systems (like CARLIY0S), and AI-generated scripts to exploit edge devices from vendors such as Cisco, Citrix, F5, and Yamaha.

Affected Systems / Environments

The identified campaigns are not limited to a single sector or manufacturer, but instead exploit attack surfaces exposed across multiple platforms:

  • Edge Devices and Network Servers: Network equipment and firewalls (such as Fortinet FortiGate), WebLogic servers, and VPN devices from Cisco, Citrix, F5, and Yamaha.
  • Data and GIS Platforms: Geographic Information Systems (GIS) servers configured with default or weak credentials.
  • AI Infrastructures and Inference Services: Exposed model execution platforms targeted to steal compute capacity and hijack AI endpoints.
  • Impacted Organizations: At least 19 organizations identified in the first analyzed campaign (16 of them in Japan, alongside targets in Taiwan and Hong Kong), in addition to global incidents impacting the banking sector in Latin America, universities in India, and industrial corporations.

Mitigation and Detection

Remediation

  • Elimination of Default Credentials: Immediately audit and modify any default credentials on GIS servers, edge devices (VPNs, firewalls), and internet-exposed management panels.
  • Hardening Endpoints and AI Services: Restrict public access to AI inference interfaces and model management platforms, enforcing strict multi-factor authentication (MFA) policies.
  • Perimeter Systems Patching: Promptly patch known vulnerabilities in network devices (Cisco, Citrix, F5, Fortinet, WebLogic) to prevent automated exploitation driven by AI-generated scripts.

Detection

  • Network and Authentication Anomaly Monitoring: Detect massive scanning patterns and automated brute-force attacks operating at speeds and volumes characteristic of AI agents.
  • Infrastructure Log Analysis: Track unusual database queries and unauthorized concurrent access to internal services following the initial compromise of a perimeter node.

“AI has increased the attackers’ workforce, but the work still consists of looking for unlocked doors. What has changed is the ability to try thousands at once, adapt to the target country’s language, and continue without fatigue.”

Wrapping Up

The deployment of artificial intelligence agents and automated frameworks like CyberStrikeAI marks a definitive transition in modern cybercrime, where offensive capabilities transition from experimental tools into permanent operational components of threat groups. Although the exploited vulnerabilities remain traditional vectors of unlocked doors and poor configurations, the industrialization and speed provided by AI require defensive teams to adopt equally automated response strategies and rigorous attack surface hardening.

References