AI penetration testing breakdown: How Baidu's offense-defense champion ARTEX was alleged to be used to breach South Korean banks

CrowdStrike's October 7 report stated that multiple South Korean financial institutions were breached from late September to early October, with attackers using the open-source AI autonomous penetration system ARTEX, and the original author Autumn-27 subsequently turned the project closed-source on October 8. ARTEX is an LLM multi-agent-driven autonomous penetration system that organizes tasks using two graphs, an asset graph and an exploration graph; the planner dispatches intents over multiple rounds through a shared to-do list, and workers can retrieve each other's execution processes. The article cited Zscaler's October 14 analysis, pointing out that its built-in default blacklist only blocks government and university domains and does not cover financial institutions, with the scope limitation written in configuration rather than in the model. Zscaler said the personal information of about 67,000 people was leaked, CrowdStrike said at least seven South Korean financial companies were affected, the attack chain included reconnaissance, SQL injection and authentication bypass, data exfiltration and monetization planning, and the main model was DeepSeek v4.1-flash.