One Person With AI Breached 90 Organizations in South Korea

South Korean police say a 32-year-old former penetration tester used an autonomous AI framework called ARTEX AI to breach more than 90 organizations and steal about 8.2 million records since late August. The campaign hit schools, hospitals, corporations and golf courses. In 38 incidents the group demanded 360 million won, about $270,000, to not leak data. Police say what once required a skilled group can now be done by one person.

The details here come from South Korean police disclosures reported 28 October 2026 via The Digital Pulse, and from related reporting on CyberStrikeAI with 44 agents against Japan targets and the earlier ARTEX campaign that hit 100 companies for $5,000.

What did the police say happened?

The police disclosure on 28 October follows the Reuters investigation on 19 October that showed a separate Chinese speaking operator using AI agents via exposed Cairn dashboards to breach more than 100 company websites for about $5,000 in tokens. The South Korean case is distinct because it includes an arrest narrative and a direct quote from investigators that the framework, not the operator's manual skill, carried the breadth.

According to The Digital Pulse, the South Korean campaign stole an estimated 8.2 million records and demanded payment in 38 incidents. A parallel Japanese case used CyberStrikeAI against 16 entities.

How does ARTEX AI change the math for attackers?

ARTEX AI is described as an open source autonomous penetration testing framework that can run an entire attack chain with limited human input. The operator supplies a target and a goal, and the framework decomposes the goal into reconnaissance, vulnerability search, exploitation, and post exploitation, with agents handling each step.

In the South Korean case, the alleged leader already understood the phases. The change is that one person could run 90 parallel intrusions without a team. The $270,000 across 38 incidents averages about $7,100 per extortion, which suggests volume over a single ransom.

I ran ARTEX style harnesses in isolated microVMs at Krova Cloud. The bottleneck that disappears is persistence across 90 targets at once. An agent framework keeps a separate context per target and retries for hours, which matches the 23 hour persistence seen in the earlier 100 company campaign.

How does this compare to the earlier 100 company campaign?

The two campaigns share tooling and differ in outcome.

• 100 company campaign (Reuters, 19 Oct): Chinese speaking operator, Cairn, 1,200+ e-commerce sites from 25 September, 100+ successes, about 23 hours per success, about $5,000 total, exposed dashboards found 2 October. No extortion, goal was lasting access.

• 90 organization campaign (South Korea, 28 Oct): Seven suspects, former pentester, ARTEX AI from late August, 90+ organizations, 8.2 million records, 38 extortion demands for 360 million won.

The earlier case was breadth without ransom. The Korean case is breadth with monetization, which made attribution possible. A third Japanese CyberStrikeAI case used 44 agents against 16 entities, according to The Hacker News.

What should defenders check this week?

The immediate hardening is not AI specific. It is the same hardening that would have blocked these intrusions without AI: parameterized queries, safe file upload handling, no unsafe deserialization, and no exposed control panels.

The difference is the search rate. When agents retry the same weakness for hours per target, a single unpatched edge device gets found. I treat any agent framework that can browse and write files as an untrusted user session and run those workloads in disposable microVMs at Krova Cloud, as described in agent sandbox checklist and agent credentials with least privilege.

What is the verdict and what remains uncertain?

According to South Korean police via The Digital Pulse, seven suspects have been identified and the ARTEX AI link is explicit. According to the same source, 8.2 million records were stolen and 38 extortion demands were made. The pattern matches the earlier 100 company campaign's economics, where about $50 per successful intrusion made failure affordable.

The verdict is that ARTEX style autonomy makes breadth attacks economically viable for a single skilled operator. The operator did not write 90 hand tuned exploits. He adopted a framework in late August and let agents iterate.

The limitations are threefold. First, the 8.2 million record count and the 360 million won total are police estimates, not a forensic audit. However, the 32-year-old's precise role as leader is alleged, not adjudicated, and cannot be treated as proven. Second, the Japanese CyberStrikeAI case with 44 agents is a separate cluster, so the three campaigns should not be read as a single actor. Third, the police data covers only reported breaches, but unreported attempts may exist, and the framework's success rate cannot be inferred from arrests alone.

FAQ

Is ARTEX AI a hacking tool?

ARTEX is described as an open source autonomous penetration testing framework. Like other offensive tooling, it can be used for authorized testing or for unauthorized access, depending on the operator.

Does this affect only South Korea and Japan?

No. The earlier 100 company campaign targeted e-commerce sites in the United States, Canada, Australia and Japan. The framework is not geographically limited.

What should teams audit first?

Search logs for long running probes that iterate the same set of weaknesses for hours per target, and audit any internet facing control panels or admin interfaces. This persistence pattern is similar to what I saw testing prompt injection as an architectural problem and when reviewing debugging an agent that wont reproduce.

Published on debugly.dev by Rohit Bhadani | Founder of Krova Cloud

Published on debugly.dev by Rohit Bhadani | Founder of Krova Cloud