Ledger / Sep 2026
Operator uses ARTEX and multiple models to breach South Korean financial institutions
Confirmed
CrowdStrike examined exposed ARTEX configurations, Claude Code session histories and memory files documenting operational AI use.
- Model
- DeepSeek, GLM-5.3, Grok 4.6, Claude CodeDeepSeek · open-weight; Z.ai · open-weight; xAI · openness unknown; Anthropic · closed
- Agent harness
- ARTEX, Claude Codeopen-source harness
- Attributed country
- Chinalanguage or infrastructure clues only
- Sectors hit
- financeSouth Korean financial institutions, including Shinhan Bank-related loan-broker…
The campaign ran from late September into early October. Flare's account of CrowdStrike's findings identifies DeepSeek as the primary model, with other models supplementing Claude Code sessions. The operator's identity and relationship to named adversaries remained unestablished in that investigation.
Sources
- AI Models are Fueling a New Wave of Cybercrime: A Quantitative AnalysisFlare · 20 Oct 2026
- China APT Uses AI Pentest Tool Artex in ConnectWise AttacksSecurityWeek · 23 Oct 2026
- Date
- Sep 2026
- Disclosed
- 7 Oct 2026
- Actor
- Unnamed Chinese-speaking, financially motivated operator (criminal), as attributed by Flare
- Targets
- South Korean financial institutions, including Shinhan Bank-related loan-broker infrastructure
- What the AI did
- Agentic intrusion
- Model as named
- DeepSeek v4.1-flash, GLM-5.3, Grok 4.6, Claude Code
- Model basis
- stated by researcher
- Autonomy
- semi-autonomous
- Outcome
- Several financial institutions breached; Shinhan Bank reported exposure of approximately 25,000 customers' data through a loan-broker application-checking service.