South Korea suspects AI agents in bank breaches that leaked data on at least 25,000 Shinhan customers

Add up the counts The New York Times reported, as relayed by Tom's Hardware, and the known total is 25,208 people ^[16]. Shinhan's 25,000 are about 99.2 percent of that ^[17], and its leak covered personal credit information ^[1]. Kookmin and Hana together account for 208 people ^[18]. Twenty of those are current or former Kookmin employees, so that leak reached staff as well as customers ^[14].

Tom's Hardware describes the investigation as one into attacks using AI agents ^[2]. So far the public evidence for that is the president's own account. "In some hacking incidents, signs have emerged of AI being used, causing considerable public concern and anxiety," Lee said at Tuesday's cabinet meeting ^[5]. "Please establish the circumstances swiftly and clearly, and concentrate personnel and resources on minimising the damage," he said ^[6]. In the same livestreamed meeting he added: "It's now become possible to use AI to hack with ease even without specialized skills." ^[7] Authorities have not disclosed which AI tools were used or the full scale of the breaches ^[9]. Tom's Hardware reported no sign yet of who is behind the attacks ^[10].

Those are two different claims. The first concerns these incidents. The second is a general statement about attackers, and it is the one a defender would plan around. For it to change a bank's threat model, investigators would have to show two things. First, a general-purpose agent did reconnaissance or exploitation that its operator could not have done unaided. Second, the weakness it used also exists at other banks. Tom's Hardware's own assessment is narrower: AI models have not replaced skilled attackers yet, though they have made planning, reconnaissance and attack execution easier ^[11].

Which model was used decides where the rest of the evidence sits. If it was a commercial agent, part of the record is with the provider, and providers can be slow to report. In September, Australia said an OpenAI agent had breached a government health data portal in June and gained unauthorised access to files ^[12]. The Australian government complained that OpenAI took 84 days to tell it ^[13]. If the attackers ran an open-weight model on their own hardware, there is no provider to ask. Many such models also lack the guardrails the large labs are believed to apply, according to Tom's Hardware ^[19].