AI in action · Investor protection

An Off-the-Shelf AI Coding Agent Was Used to Break Into South Korean Banks. The Attacker Left the Chat Logs Behind

CrowdStrike reported on October 7, 2026 that intrusions at South Korean lenders between late September and early October were carried out with ARTEX, an open-source, Chinese-developed penetration-testing tool, driven through a commercial AI coding agent. The attacker's exposed server held the agent's session histories, configuration files and a CV. At least five banks, including Shinhan and KB Kookmin, were affected; about 68,000 records were reported exposed.

Oct 19, 2026

The first documented use of an AI coding agent to attack banks did not come from a state or a criminal syndicate. According to a report CrowdStrike published on October 7, 2026, the intrusions at South Korean financial institutions between late September and early October were the work of a single, probably Chinese-speaking and financially motivated actor, identified with what the company called moderate confidence, who used ARTEX, an open-source penetration-testing framework developed in China, and drove it through Anthropic's Claude Code, a commercial AI coding agent. ARTEX can connect to several large language models; in this case its main backend was DeepSeek, with other models used in separate sessions.

The evidence came from the attacker's own carelessness. CrowdStrike found open directories on the attack infrastructure containing the coding agent's session histories, ARTEX configuration files, the agent's memory files, a Chinese-language prompt instructing it to behave as a penetration tester, and what appeared to be the operator's CV. The Register reported the CV detail; several outlets reported that the logs point to a 26-year-old in Guangdong, an identification that remains unconfirmed. The picture is of one person directing an agent that wrote and ran the attack code, rather than a team of specialists.

The scope is disputed only at the margin. The Korea Times reported at least five lenders affected, including Shinhan Bank and KB Kookmin Bank; the Korea JoongAng Daily counted seven banks and lenders; other reports put the number at nine institutions. About 68,000 customer records were reported exposed in total. No customer financial losses or stolen passwords have been confirmed, and the security firm Reflectiz noted that which of the affected banks were hit by ARTEX specifically is partly inference from the recovered logs.

The episode has already changed what supervisors ask. Reuters reported on October 13 that the Federal Reserve, the OCC and the FDIC are questioning US banks about whether they use Chinese-developed technology and whether they deploy AI agents in cyber defence, questions that track this case directly. The European Systemic Risk Board warned in July that frontier models could compress the time between an exploit's discovery and its automated use from weeks to hours; Christine Lagarde noted in October that in a simulated 32-step attack the latest models completed every step where models from late 2025 had managed a third. A lone operator with a free tool and a commercial agent breaching several banks in two weeks is what that warning looks like in practice.

For anyone with a bank account the practical reading is dull but real: the attack surface of a bank is now also the attack surface of the AI tools its adversaries can rent by the hour. For investors in the AI companies, the case is the first of what will be many tests of whether model providers' safeguards hold up when the user is openly running a penetration-testing framework against a bank. CrowdStrike's report does not say whether any of the providers detected the activity.

Sources: Infosecurity Magazine — Chinese Hacker Deployed AI in Campaign Against South Korean Banks (Oct 2026); The Register — CrowdStrike finds possible bank hacker's CV among exposed AI logs (Oct 8, 2026); Security Affairs — AI-driven tool ARTEX used in attacks against South Korean banks (Oct 2026)