A cyber campaign targeting South Korean financial institutions between late September and early October 2026 used artificial intelligence tools, according to cybersecurity firm CrowdStrike. Investigators trace the activity to a 26-year-old suspect in Guangdong province, China, who allegedly utilized an open-source testing platform alongside Anthropic’s Claude.
South Korean financial institutions faced a wave of digital intrusions as a string of cyberattacks hit at least nine banks starting in late September. The scale of the breach prompted an official response. South Korean police launched a probe into the incidents, while President Lee Jae Myung called for robust response measures across the financial sector.
Individual institutions began disclosing the extent of the data exposure as authorities stepped up their probe. Shinhan Bank reported that personal information belonging to about 25,000 customers had been compromised. Meanwhile, KB Kookmin Bank confirmed that the personal data of 119 of its customers was leaked during the campaign.
Claude Code Sessions Reveal Attacker Details in Guangdong
The investigation into the financial sector breaches advanced when researchers analyzed coding-tool sessions and digital infrastructure tied to the attacks. CrowdStrike published a report detailing personal details uncovered within the attacker’s workflow. Prompts entered into Anthropic’s Claude Code assistant revealed specific personal identifiers, including a Telegram account, an age of 26, educational background, and a location in Maoming, a city in China’s Guangdong province.
Security analysts found that the user asked the AI assistant to draft a security researcher resume incorporating those personal details and summarizing results from the hacking activity. Additional prompts asked the model where threat actors typically sell Korean data breach information and requested help locating Telegram groups dedicated to data sales. When contacted by researchers via a phone number found in the logs, a man who answered the call stated he had no knowledge of the matter.
ARTEX and Large Language Models in Automated Testing
The technical architecture of the attacks relied on a recently published Chinese-developed utility. The attacker utilized ARTEX, an open-source AI agent designed for automated penetration testing. Published on GitHub in 2026 by a security engineer using the handle Autumn, the software functions as a bridge rather than a standalone model, connecting to external large language models such as ChatGPT, Claude, and DeepSeek.
While the utility’s public repository states it is intended for personal learning, code research, and local technical verification—and warns against targeting online systems—analysts observed it being deployed in the wild. CrowdStrike assessed that the operator was likely a Chinese speaker driven by financial motives, though the firm noted the assessment carried moderate confidence because the activity had not been attributed to a named adversary.
“This assessment is made with moderate confidence based on the use of the Chinese-developed tool ARTEX and observed Chinese-language prompts.”
CrowdStrike
Broader Implications for Automated Cyber Defense
The deployment of autonomous testing tools against banking infrastructure highlights growing security challenges surrounding artificial intelligence agents. The incident in South Korea follows a similar international event from earlier in the year. Australia disclosed in September that an OpenAI autonomous agent breached a government health statistics portal in June, marking one of the first known instances of an AI agent hacking a government system.
As police investigations continue in Seoul, representatives for Anthropic, the South Korean police, and China’s Foreign Ministry did not immediately respond to requests for comment regarding the CrowdStrike findings.