Seoul police allege a former pentester used AI on 90 firms; the referral covers 15 firms and 11% of 8.2 million records

The referral covers 15 firms and 11 per cent of the 8.2 million records

The Seoul Metropolitan Police Agency's release, read on 27 October 2026, says the suspected leader was referred to prosecutors in custody on 23 October for 15 victim companies and about 900,000 personal records. The same release says the case covers at least 90 companies and about 8.2 million records. So the referral covers 17 per cent of the companies and 11 per cent of the records (both derived). The other 75 companies, which the release says have not yet reported, and about 7.3 million records, or 89 per cent (derived), are to be referred after further investigation.

That is the fact under the headline that spread on 26 October, "Former pentester used ARTEX AI to steal 8.2 million data records" (BleepingComputer, 21:08 UTC). Everything below is an allegation. The police say the suspected leader is in custody and six others are being investigated without detention. The sources we read state no indictment, plea or conviction. The release gives the 8.2 million as drawn from digital forensic analysis and the suspected leader's own statement, and says the stolen data was recovered. It does not name ARTEX. The name appears in the BleepingComputer text, and in neither the police release nor the Yonhap report that we read.

How this was read: the primary is the release on the Seoul police website, in Korean, which we read in full. The translations here are ours. Yonhap's report of 26 October and BleepingComputer's report of the same evening are secondary. We read each host's robots.txt before requesting a page from it. All three texts were read on 27 October 2026, between 08:05 and 08:40 UTC.

Zoom

What the release states, and what it leaves out

The release is a press statement, not a court document. It prints the legal provisions the police used and the dates of the arrests, and it describes the method in three sentences. It gives no case number, court, indictment or victim name. The suspected leader is identified only as A. Table 1 sets what the release says against what it does not.

"Former pentester" is a label from the release's background line, and it explains nothing about access

The release says the suspected leader is not an IT graduate, learned hacking alone from blogs, and had a record of performing mock hacking work for security checks. Yonhap, reporting the same announcement, puts that work at a security firm from 2019 to 2022. BleepingComputer turns that into "former pentester" and "background in penetration testing".

What the record does not establish is any link between that job and the 90 companies. It does not say a victim was a client, that a tool, credential or scope document from the job was reused, or that the former employer did anything wrong. The release's own sentence runs the other way: self taught, not an IT degree. The headline invites a reader to think insider knowledge made the difference. The primary gives no basis for that, so treat it as an inference nobody has tested.

There is a friendly name at work, and it is worth taking apart. "Penetration test" is what the activity is called when someone has a contract. The same steps without one are, in the statute the police cite, network intrusion. A tool described as an autonomous penetration test system inherits the respectable word, and its description says what it can do, not who may point it at what. Our briefing on the CrowdStrike report recorded that ARTEX's own project page restricts its use. A line in a licence stopped nothing here, if the tool was involved at all.

Is it ARTEX, and is it the bank tool? The name sits in one text

The Korean police text says the suspect used, among other things, an "AI autonomous penetration test system" that was "recently used" in personal information leaks from the financial sector (release, our translation). Yonhap repeats that wording. Neither names a product. BleepingComputer writes "ARTEX" and says it is "the same autonomous framework used in the attacks against" South Korean banks. BleepingComputer says the Seoul police told Yonhap that the suspect used it. The word ARTEX is not in the release or in the Yonhap article we saved, which describes the system without a name. So on what we could read the name is BleepingComputer's, and we could not trace it to a police statement.

Zoom

On the earlier briefings: brief 223 found AI use in the bank breaches unconfirmed on the record, and the faults officials described were old ones. Brief 270 read CrowdStrike's report of 7 October, which names ARTEX from the actor's own files and ties it to the banks through press reports, with no bank named. Nothing in the new texts settles either. If the Korean police mean the same open source tool, this is a second user of it, and nothing in the texts links the two users. CrowdStrike's language call about the bank actor was moderate confidence and rested partly on tool choice. The same kind of tool appearing in a second case, with a suspect the release does not tie to the bank actor, is a reminder that tool choice describes the tool and not the operator (our inference, conditional on the tool being the same, which the police texts do not say).

How much the AI did: the only number is a count, and the police have an interest

The release says the suspected leader first operated a website flaw finder directly and then, from late August 2026 according to Yonhap, used generative AI to hit more companies without discrimination. Yonhap adds that the police count 60 of the victim companies as hacked using AI. The release does not carry that figure.

That count allows one piece of arithmetic, and it is rough. If all 60 AI victims fell between 25 August (our reading of late August) and the 15 October arrest, about 7.3 weeks, that is about 8 a week. The other 30 over the earlier 51 weeks is about 0.6 a week, so about 14 times slower (all derived, from "about 90" and "late August"). That fits the police's point that AI widened the attack. It does not show the AI caused the widening. The release gives no intrusion dates per victim, so a tool change, a second operator, a larger target list or the accomplices' phone threats could produce the same pace. The release also gives no prompt, model, vendor, error rate or comparison with the earlier tool.

Zoom

Two more points. First, the police's closing line says an individual, not a professional hacker group, can now run large scale hacking with AI. The release's own account has seven suspects, a phone and text team of five, and a Telegram instigator, so the operation was not one person. Second, the Seoul agency's interest: the same release says the national investigation office provided it with a high performance AI computer, that it built three internal AI analysis tools, and that all seven suspects were caught within four months of the case opening. It ends with a plan to extend AI in investigations. An agency announcing both the AI threat and its own AI answer is not wrong to do so, but the sentence about AI lowering the barrier is the agency's judgement, listed in the release under implications, not a finding.

Extortion: 38 firms, 360 million won, and no payment on the record

The police say 38 of the roughly 90 companies, 42 per cent (derived), were sent threats, and that 360 million won was demanded in total, about 9.5 million won a firm on average (derived). BleepingComputer converts that to $270,000. That conversion is BleepingComputer's, not the police's. The release says amounts and further targets are still being investigated, and that the police are checking whether firms were threatened or paid and did not report, for fear of a leak.

The threat's wording is the release's paraphrase: comply or the information is leaked. It adds that the messages pressed victims with three consequences: damage to the company's image, a fine after an official investigation, and damages suits from customers. Those are real exposures for a firm that holds personal data, which is why they work. The police's advice is plain: do not negotiate with the offender, do not send money, report to the police at once.

Technical details

The release prints no indicators. It has no hash, domain, address, tool name, command, prompt, server log line or filename. The three tool descriptions below, the two statutes and the two extortion lines are everything technical it prints, copied from the saved release exactly. They help a defender decide what to ask a tester or hunt for, not how to copy anything.

First, the statutes. The first is the network intrusion offence and the second is extortion. The release prints the ceiling after each, which we read as up to 5 years and up to 50 million won, and up to 10 years and up to 20 million won. Those are statutory ceilings, not a forecast of any sentence.

정보통신망법 제71조 제1항 제11,14호(정보통신망침해등) ·············· 5년 이하, 5천만원 이하
형법 제350조 제1항, 제352조(공갈, 공갈미수) ···························· 10년 이하, 2천만원 이하

Second, the method lines. The first says that early on the suspected leader operated an automated tool that finds website flaws directly, then used generative AI to attack a wider range of companies without discrimination. The second says a system called an AI autonomous penetration test system, recently used in financial sector leaks, was also confirmed among the tools. The release also says a search of the suspected leader's home found the computer, hacking tools and extortion accounts.

범행 초기에는 웹사이트 취약점을 찾아 주는 자동화 도구를 직접 조작해 해킹하다가 생성형 AI를 이용해서 더 넓은 범위의 업체를 무차별적으로 공격하기 시작함
(AI 이용 해킹) 최근 금융권 대상 개인정보 유출에 사용된 ‘AI 자율 침투 테스트 시스템’ 등을 활용하여 해킹을 수행한 사실도 확인하였음
A의 주거지에 대한 압수수색을 통하여 해킹에 사용된 컴퓨터와 해킹 도구 및 협박용 계정 등 주요증거를 확보하였음

Third, the extortion lines. The first says the offenders emailed victims a message to the effect that the information would be leaked if they did not comply. The second says they pressed victims with the reputational damage, the fine and the customer damages suits. The police published a reconstruction of one email as an image through Yonhap, which carries a notice against resale, so it is not reproduced here and we did not read its text.

(협박방법) 피의자들은 해킹 이후 피해업체 측에 ‘협박에 응하지 않으면 정보를 유출하겠다’는 취지의 협박 이메일을 보냈음
특히 정보유출이 외부에 알려지면 기업의 이미지 실추와 당국 조사에 따른 과징금 부과 및 고객들의 손해배상 소송까지 이어질 수 있다며 요구에 응하라고 압박하였음

What to hunt, which is our guidance and not the release's: the release says the police traced the suspected leader from the victims' server logs, so web server and application logs are the place. Look for one source making broad automated requests across many pages and parameters of public sites, especially outside an agreed test window, and for a burst of requests that all fail in the same way and then succeed on one. Keep the logs for long enough to be useful after a threat arrives. The release does not say which flaw any victim had, so there is no vulnerability to patch from it.

For organisations that buy penetration tests, in the order worth doing

None of this describes how the offence was done, and no regulator has issued it. The order is our judgement. The release does not say any victim was a client of the suspected leader's former employer, so the first four items are about the tester relationship in general and not a claim about this case.

The question that exposes the gap

The record is a referral for 15 firms, a headline for 90, and a tool name that two Korean texts leave out. Set the AI label and the tester label aside and ask what the police ask of every victim: if an automated tool with no history of working for you started probing your public websites tonight, would your logs show it as a stranger, and would you hear about it from them or from an extortion email?

Key facts