South Korea Investigates First Known AI Agent Cyberattacks on Commercial Banks

On Tuesday, October 6, 2026, South Korea's National Police Agency launched a formal investigation into a coordinated cyberattack that breached at least seven domestic financial institutions. The intrusions exposed the personal data of approximately 68,000 individuals. Investigators suspect the attackers deployed an autonomous artificial intelligence agent to execute the breaches, marking the first known instance of AI-driven tools successfully penetrating the global financial system.[1][2][3]

The attacks began in late September and targeted secondary infrastructure rather than core banking ledgers. Hackers breached employee portals, loan applications, and sales-support systems. Shinhan Bank, one of the country's largest lenders, reported that the data of about 25,000 customers was exposed after attackers bypassed a verification step on a lookup service built for external loan recruiters.[2][4]

Other major institutions, including KB Kookmin Bank and Hana Bank, also suffered data leaks. The compromised information included sensitive financial details such as annual incomes, personal loan limits, names, and phone numbers. The Financial Services Commission immediately convened an emergency session, instructing banks, insurers, and securities firms to fortify their externally accessible systems.[3][4]

The mechanism behind the breach

The investigation centers on ARTEX AI, an open-source penetration-testing agent developed by a Chinese cybersecurity engineer known as Autumn. ARTEX is not a standalone foundational model. Instead, it acts as an orchestration layer that connects to existing large language models—including Anthropic's Claude Opus, OpenAI's ChatGPT, and China's DeepSeek—to automate the discovery of network vulnerabilities.[2]

Designed to help organizations stress-test their own defenses, ARTEX can autonomously scan systems, write exploit code, and analyze stolen data with minimal human oversight. Security researchers found traces of the tool, including a specific Chinese-language string, on the compromised South Korean web servers. The attackers effectively weaponized a defensive diagnostic tool to automate credential stuffing.[2]

By delegating the repetitive labor of probing side doors to an AI agent, the hackers operated at a scale and speed that traditional manual methods rarely achieve. Over a 30-hour window starting September 28, the automated system rapidly fed random customer numbers into Shinhan Bank's external portals. This continuous, machine-driven persistence allowed the attackers to slip past standard verification hurdles.[2]

"It's now become possible to use AI to hack with ease even without specialized skills," South Korean President Lee Jae Myung said during a televised cabinet meeting on Tuesday. He directed his administration to overhaul the nation's security paradigm for the AI era and deploy personnel to minimize the damage. "Signs have emerged" that AI models were used, he added, "causing considerable public concern and anxiety."[1][3]

Tracing the automated intrusion

The National Office of Investigation has assigned 28 investigators from its cyberterrorism unit to track the perpetrators. However, the autonomous nature of the attack complicates the digital forensics. The intrusions were routed through more than 30 internet protocol addresses spread across roughly 12 countries, including the United States, Japan, and Germany.[3]

Because hackers frequently spoof their locations to mask their origins, authorities have not yet identified the individuals or groups responsible. The South Korean government is currently seeking international cooperation to trace the network traffic back to its source. It remains unclear whether a state-sponsored entity or an independent criminal syndicate orchestrated the campaign.[2]

Following the public disclosure of the breaches, the developer behind ARTEX AI updated the software's user guidelines to explicitly prohibit unauthorized intrusions and data theft. Yet cybersecurity experts note that policy updates offer little practical defense. Because the tool is open-source, anyone can download the underlying code, strip away the usage restrictions, and deploy the agent maliciously.[2]

A shifting security landscape

The South Korean breaches highlight a fundamental inversion in cybersecurity economics. The same capabilities that allow a security team to efficiently map their own vulnerabilities now allow attackers to exploit those gaps at a fraction of the historical cost. Financial institutions are finding their peripheral applications uniquely vulnerable to automated probing.[1]

While banks have spent decades hardening their primary mainframes, modern banking relies on a vast web of application programming interfaces, third-party vendor connections, and remote employee portals. AI agents excel at mapping these complex, interconnected surfaces to find the single forgotten endpoint that grants access to sensitive data.[2]

The Financial Supervisory Service and the Financial Security Institute have begun sharing the identified malicious IP addresses with the broader financial sector to block further automated traffic. They are also auditing how external contractors, such as the loan recruiters targeted in the Shinhan Bank breach, are granted access to sensitive internal databases.[3]

Critics have argued that giving non-employees direct lookup access to customer credit data represents a severe lapse in basic cyber hygiene. The AI agent did not invent a complex zero-day exploit; it merely possessed the tireless capacity to find and abuse an existing administrative oversight. This distinction is crucial for understanding how agentic AI changes the threat landscape.[2]

The broader context of AI-assisted hacking has accelerated throughout 2026. Earlier in the year, Anthropic disclosed that state-sponsored Chinese hackers had utilized its AI technology to automate break-ins at approximately 30 targets worldwide. Similarly, Australian officials recently reported that an OpenAI agent had infiltrated a government healthcare-statistics portal, accessing both public and nonpublic files.[1]

These incidents demonstrate that foundational models, despite their built-in safety guardrails, can be orchestrated by secondary agents like ARTEX to perform malicious tasks. The agent breaks a complex intrusion into smaller, benign-looking queries that the foundational model answers without triggering safety filters. The agent then pieces those answers together to execute the attack.[2]

South Korean cybersecurity stocks surged by as much as 30 percent on Tuesday as investors anticipated a massive wave of defensive spending by the financial sector. Banks are now expected to deploy counter-agents—AI systems specifically trained to detect the behavioral patterns of offensive AI tools, such as the rapid, non-human cadence of credential stuffing.[3]

Regulators are now urging financial leadership to adopt a hacker's mindset, using their own AI agents to continuously stress-test their external perimeters. As the investigation continues, the immediate focus remains on securing the compromised accounts and determining whether the stolen financial profiles have already been sold or utilized on the dark web.[2]

The 68,000 compromised accounts represent a relatively small fraction of South Korea's highly digitized banking population. However, the methodology guarantees that this will not be an isolated event. The barrier to entry for launching a coordinated, multi-institution cyberattack has permanently fallen, shifting the burden of speed entirely onto the defenders.[1]