ARTEX AI-Driven Attack Campaign Hits Seven South Korean Financial Firms

Summary

An unknown, likely financially motivated threat actor used the ARTEX AI-driven autonomous penetration testing system to compromise several South Korean financial institutions. The attacker leveraged Large Language Models (LLMs), including DeepSeek and Claude Code, to accelerate vulnerability discovery and exploitation. The campaign resulted in the theft of roughly 67,000 records containing customer income, loan, and other sensitive financial data.

Investigation

CrowdStrike investigators traced the activity after discovering an unsecured open directory hosted on the attacker’s server. The directory exposed session logs, configuration files, and memory artifacts that revealed the attacker’s methodology and identifying characteristics. Researchers connected the operation to the ARTEX framework and identified a network of proxy IP addresses used to access external LLM APIs.

Mitigation

Organizations should broaden asset inventories to include edge systems, third-party integrations, and mobile support platforms. Continuous automated red-teaming with local, unconstrained AI models can help uncover weaknesses more frequently than traditional annual assessments. Securing all internet-facing endpoints against Insecure Direct Object References (IDOR) and similar access-control flaws is also essential.

Response

When suspicious queries or unauthorized access to customer data endpoints are detected, organizations should immediately review API logs and active session tokens. Enforce strict rate limiting on customer ID lookup services to reduce automated enumeration. Security teams should also hunt for unauthorized agent-based tools and unusual traffic patterns associated with known LLM proxy infrastructure.

Simulation Execution

Prerequisite: The Telemetry & Baseline Pre-flight Check must have passed.

Rationale: This section details the precise execution of the adversary technique (TTP) designed to trigger the detection rule. The commands and narrative MUST directly reflect the TTPs identified and aim to generate the exact telemetry expected by the detection logic. Abstract or unrelated examples will lead to misdiagnosis.

• Attack Narrative & Commands: An adversary has gained initial access to a developer workstation. To maintain consistency in their automated exploitation framework, they attempt to drop a CLAUDE.md file in the user’s temporary directory. This file contains specific instructions in Chinese to guide an AI-driven agent through further reconnaissance and lateral movement. The goal is to see if the SOC detects the presence of this specific configuration file.

• Regression Test Script:

  # Simulation script to trigger the detection rule
  $targetPath = "$env:TEMPCLAUDE.md"
  $content = "Instructions for AI Agent: Perform reconnaissance on local network."
  
  Write-Host "[+] Creating sensitive file: $targetPath"
  Set-Content -Path $targetPath -Value $content
  
  Write-Host "[+] File created. Check SIEM for detection."

• Cleanup Commands:

  # Cleanup the simulation artifacts
  Remove-Item -Path "$env:TEMPCLAUDE.md" -Force -ErrorAction SilentlyContinue
  Write-Host "[+] Cleanup complete."