The Hacker Asked Claude to Write His Résumé. South Korea Still Can’t Agree Who Investigates Him.

A person seen only in silhouette, lit by a laptop screen, working alone at night, illustrating the solitary hacker who asked an AI tool to draft his resume.

An AI hacking investigation now spans seven South Korean financial institutions, after CrowdStrike published a report on 7 October linking a monthslong intrusion campaign to a Chinese-speaking, financially motivated hacker working alone. Other reporting puts the tally as high as nine, including Shinhan Bank, KB Kookmin Bank, Hana Bank and BNK Financial Group. Shinhan alone lost personal data on roughly 25,000 customers through a loan broker inquiry service. Investigators now estimate close to 68,000 people affected across the full campaign, a figure the publicly confirmed breaches at named banks do not come close to explaining on their own.

 

A Résumé Written by the Tool That Caught Him

CrowdStrike’s analysts recovered open directories containing Claude Code session histories, memory files and configuration data tied to ARTEX, an open-source Chinese penetration-testing agent the attacker ran on DeepSeek’s v4.1-flash model. In separate sessions, he turned to Claude Code directly, asking it to draft him a security-researcher résumé. What came back included a Telegram handle, an age, an education entry at South China University of Technology, and a location in Maoming, Guangdong. He also asked Claude where stolen Korean data is typically resold and how to find the Telegram groups that trade it. CrowdStrike’s own language is carefully hedged: the attacker is likely a Chinese speaker and financially motivated, assessed with moderate confidence. The résumé details probably belong to him. Probably is the honest word, and CrowdStrike used it.

Adam Meyers, the firm’s senior vice president of counter adversary operations, framed the significance plainly. This is “an example of a human adversary leveraging AI agents to conduct widespread attacks,” he said, one that “allows one human to target many customers in a very short period of time.” The tools did not make the attacker smarter. They let one person do the work of a team, and in doing so, they left behind a different kind of evidence trail than a team usually does, a searchable record of exactly what he asked, in his own words, timestamped.

 

The AI Hacking Investigation Nobody Has Agreed to Own

South Korean police escalated their inquiry into a formal investigation on 6 October, assembling a 28-member task force under the Information and Communications Network Act. On 21 October, they referred the case up to the Seoul Metropolitan Serious Crimes Investigation Agency, arguing that suspected breaches of the Electronic Financial Transactions Act qualified it as a serious crime under the governing act. Two days later, on 23 October, that agency declined to open an investigation. Public reporting does not say why, and it does not say who picks the case up next either.

That sequence deserves more attention than the résumé detail everyone is quoting. A months-long, AI-assisted campaign against at least seven financial institutions, with an attribution researchers are comfortable naming at moderate confidence, still could not settle a basic governance question: whose job is this. The evidence existed in full. The ownership did not.

 

The Parallel Most Organisations Haven’t Tested

Most organisations now log some version of what their own people do inside AI tools. Far fewer have assigned anyone the job of reviewing those logs for the kind of anomaly this case turned on, a prompt that reads nothing like ordinary work. A request to draft a cover résumé, sitting next to questions about where to sell stolen data, is not a subtle signal once somebody looks. It is only useful if somebody is positioned to look, and empowered to act on what they find there.

South Korea’s police force and its own Serious Crimes Investigation Agency are two separate, well-resourced institutions, and they still could not agree on ownership within two days of the referral landing. Most PMOs have never had that conversation once about their own AI tool logs, let alone tested whether the answer they would give actually holds up under pressure.

 

What This Leaves for Every Governance Review

Pick the AI tool in your own environment that generates the most usage logs nobody has reviewed end to end this year. Name who would own investigating an anomaly inside it today, and ask whether that answer would survive being tested the way South Korea’s just was.

Get new articles by email. One short digest a week, no spam, unsubscribe any time.

Lighting Up The Tunnel, in your inbox

Pre-Mortems, cheat sheets and practical ideas on PMO, governance, AI and cyber for leaders. Short, useful and free.

No spam. Unsubscribe any time. Read our privacy policy for more info.

Scroll to Top