12 OCT 2026 — In late September, an attacker stole customer data from a series of South Korean banks. When CrowdStrike examined the attacker's own exposed servers, it found an open-source AI hacking tool built in China and driven by commercial AI models, plus session logs from Anthropic's Claude Code.
The breaches hit some of the country's largest lenders and prompted financial regulators to hold an emergency meeting on a public holiday. The files show how much ground a financially motivated operator, or a small group, could cover in a few days with AI tools in the loop.
What happened at the banks
Between 27 and 30 September, data was taken in intrusions at Shinhan Bank, KB Kookmin Bank, Hana Bank, BNK Busan Bank, Yegaram Savings Bank and Hyundai Capital, the Kyunghyang Shinmun reported, citing financial authorities. Welcome Savings Bank and two peer-to-peer lenders also reported leaks.
Shinhan lost 25,727 records, including names, contact details, annual incomes and loan limits. Kookmin and Hana lost data on 119 and 89 customers. The attacker went after weakly protected internal systems rather than customer banking apps: at Shinhan, a loan-progress inquiry service used by brokers; at Kookmin, a mobile work-support system for staff.
Regulators held an emergency inspection meeting on 4 October, a holiday, and President Lee Jae Myung called for a thorough investigation.
The tool: ARTEX
ARTEX is an "agentic" penetration-testing system that uses large language models to plan and carry out hacking steps with little human input. It was built as a red-team tool for finding weaknesses before attackers do. According to the Kyunghyang, it appeared on GitHub on 26 July, and its latest version came out on 24 September, three days before the attack on Kookmin began.
A Korean security researcher reported finding the string "ARTEX" in the page titles of web servers used in the attacks, the paper reported.
What the attacker's files showed
CrowdStrike's report, published on 7 October, traces two servers the attacker controlled. One hosted the ARTEX instance. The other, in Hong Kong, held open directories of Claude Code session histories, ARTEX configuration files and Claude memory files.
According to the report, ARTEX ran on DeepSeek's v4.1-flash model, which the attacker likely accessed through an API reseller, and the attacker added Zhipu AI's GLM-5.3 and xAI's Grok 4.6 in further Claude Code sessions. A Chinese-language prompt told the model how to conduct the penetration test. In one session, the attacker asked Claude where Korean breach data is usually sold and for help finding Korean Telegram groups that trade it.
Who was behind it
CrowdStrike has not attributed the attack to a known group. It says, with moderate confidence, that the attacker is a Chinese speaker and financially motivated. The files also contained personal details that CrowdStrike says it cannot definitively link to the attacker, and which we are not repeating.
Why it matters beyond Korea
Korean officials and experts quoted by the Kyunghyang were careful about what the tool proves. Park Sang-won, president of the Financial Security Institute, noted that ARTEX is open source and publicly available, and that the attack routed through addresses in eight countries, so neither reveals the attacker's nationality.
CrowdStrike's own conclusion is about speed. The activity "demonstrates how AI tooling can enable a financially motivated threat actor to conduct multiple intrusions within a short time span," it says, and it expects attackers to keep experimenting with such tools. One security executive quoted by the paper said the banks' core systems held up because they need more than a password. The attacker got in through peripheral systems that lacked that protection.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, whose Claude Code tool the attacker used. Readers should weigh that interest.
View author profile → · Editorial policy
