South Korean banks hacked: attacker used AI tool ARTEX and Claude Code, CrowdStrike says

U.S. cybersecurity company CrowdStrike published a report on Wednesday, October 7, on a series of intrusions against South Korean financial institutions. The attacks took place between late September and early October 2026. According to the report, the attacker relied heavily on artificial intelligence agents, notably ARTEX, a penetration-testing tool built in China, and on several large language models.

What we know

The South Korean press links this activity to several data breaches at Shinhan Bank, KB Kookmin Bank and Hana Bank. According to CrowdStrike, the compromised systems included a loan application tracking service used by brokers and an internal mobile app for the employees of another bank. Financial authorities and investigators have opened inquiries. According to The Korea Times, cited by The Register, at least five institutions were affected: about 25,000 customers at Shinhan Bank, 119 at KB Kookmin Bank and 89 at Hana Bank, plus Yegaram Savings Bank and BNK Busan Bank. According to Reuters, at least nine banks have reportedly been targeted since late September. CrowdStrike, for its part, does not confirm the total number of targeted organizations.

CrowdStrike assesses with "moderate" confidence that the attacker speaks Chinese and is financially motivated. The company does not link the attacker to any known group. Their identity has not been established: a résumé written at their request by the AI contained personal details that the researchers believe probably belong to the attacker but cannot attribute to them with certainty. For example, the age given does not match the stated date of birth.

AI agents at work

ARTEX is a so-called "agentic" penetration-testing tool: it hands task execution to AI agents connected to external language models. On servers controlled by the attacker and left open, the researchers found Claude Code session histories (Anthropic's programming assistant), ARTEX configuration files and Claude memory files. These traces made it possible to reconstruct a large part of the operations.

According to the report, the main engine of the ARTEX instance was a DeepSeek model (DeepSeek v4.1-flash). The GLM-5.3 and Grok 4.6 models were used as a complement in Claude Code sessions. The attacker also asked Claude where stolen data is usually sold in Korea and which Telegram groups trade in it.

After its malicious use was confirmed, ARTEX's developer closed the source code and stopped updates. They say the tool was meant for learning and research. Copies and derivative versions remain in circulation. Anthropic had not responded publicly at the time of writing.

Key takeaways

• Automated attacks are becoming accessible: AI tools lower the skill level needed to carry out a complete intrusion.
• Monitor services exposed to partners (brokers, vendors) and internal mobile apps, which are often less protected than the core system.
• According to the Korean press, the country's financial regulator is urging banks to analyze how requests arrive rather than who is connecting: blocking by IP address is no longer enough.
• Attackers make mistakes too: poorly protected AI logs made it possible to trace much of the operation.