AI tools accelerate attacks on South Korean banks

Several South Korean financial institutions were targeted between late September and early October 2026 in a cyber‑attack in which artificial intelligence accelerated the search for vulnerable systems. Security firm CrowdStrike reconstructed the sequence using files and session logs on a server that the attacker controlled. The Decoder and the Taiwanese specialist outlet iThome reported independently on the analysis. Both sources stress that the attribution should be read cautiously: CrowdStrike considers a Chinese‑language attacker likely, but assigns only medium confidence to this and does not link the activity to any known state‑backed or criminal group.

ARTEX automates vulnerability searching

At the centre of the incident is ARTEX, an open‑source penetration‑testing tool. It was released on GitHub in July 2026 and was developed in China. The program combines classic testing methods with large language models. Instead of manually examining each target address and possible vulnerability, an operator can use it to automate workflows, evaluate technical hints and prepare attack steps. According to CrowdStrike, ARTEX was configured to use DeepSeek v4.1‑flash as its primary language model. Additional session logs also showed GLM‑5.3 and Grok 4.6 being used as helpers.

The recovered data, according to the security researchers, include ARTEX configuration files, logs from Claude Code and saved work notes. They reveal not only technical steps but also that the attacker searched Telegram groups for marketplaces for stolen data. These traces suggest that the harvested information was intended for economic exploitation. However, they do not prove that a sale actually took place. The sources have so far reported neither confirmed buyers nor a fully documented transfer of the datasets.

At least 25,000 records at one bank

The details are especially concrete for Shinhan Bank. The Korean newspaper Khan reported that more than 25,000 records had been exfiltrated there. They included names, contact details, income and credit limits. Other banks are also said to have been affected, but their names and the exact number of stolen records were not fully disclosed in the available reports. Consequently, neither the total number of victims nor the potential financial damage can be reliably quantified. Consequences such as identity theft, fraud or abusive credit applications have not yet been quantified.

The South Korean financial regulator convened an emergency meeting after the incidents became public. President Lee Jae‑myung demanded a comprehensive investigation. These reactions demonstrate the political significance, but they do not replace a completed forensic examination. So far it has not been publicly proven whether all targeted institutions shared the same vulnerability or whether ARTEX found different entry points in each case. Information is also lacking on how long the attackers remained undetected in the networks and which security controls failed.

AI changes the pace, not the evidentiary rules

The case does not prove that a language model alone can compromise a bank. Rather, it shows how a single operator can combine existing tools, open models and traditional attack techniques. Automated searching and machine‑generated suggestions shorten the path from a known vulnerability to a functional exploit. The decisive steps, however, remain the attacker’s technical actions. Attribution must also rely on infrastructure, language, timing patterns and other indicators. The use of a particular model or a tool developed in China does not by itself prove origin or sponsor.

• CrowdStrike attributes the attacker to a Chinese‑language environment with only medium confidence
• ARTEX combined automated security testing with multiple language models
• more than 25,000 records are said to have been exfiltrated from Shinhan Bank
• complete victim and loss figures are still missing for other institutions
• a sale of the stolen data has not been publicly confirmed

For banks this translates into a dual challenge. They must patch known vulnerabilities more quickly because attackers can increasingly automate their search. At the same time they must log unusual queries, new tools and data movements in a way that an incident can later be reliably reconstructed. Classic measures such as segmentation, multi‑factor authentication, minimal privileges and monitoring of large data outflows do not lose importance because of AI. On the contrary: the faster an attack is prepared, the more critical short response times and clearly defined shutdown procedures become.

It remains unclear how extensive the campaign actually was. CrowdStrike’s analysis is based on material from the attacker’s infrastructure; an independent verification of all details by South Korean authorities or a second security firm is not present in the cited reports. A link to a state or organized group is also not established. The most reliable conclusion is therefore narrower: open AI‑assisted attack tools were used in a real intrusion against financial institutions, and at least one bank reported a significant exfiltration of sensitive customer data. Anything beyond that must be clarified by the ongoing investigation.