AI pentesting tool ARTEX used to steal data from Korean banks

Attackers used an artificial intelligence penetration testing tool called ARTEX to break into South Korean financial organisations and steal data, according to details disclosed on 8 October 2026 by CrowdStrike Intelligence and reported by The Hacker News.

The campaign was active from late September to early October 2026, CrowdStrike Intelligence said. The researchers described it as targeted activity aimed at financial organisations in South Korea, and said it resulted in data exfiltration — that is, data being copied out of the victims' systems. The number of organisations affected, the volume of data taken and the identity of the victims were not stated in the available reporting.

ARTEX is described as an AI penetration testing tool. Penetration testing tools are built for defenders and hired testers: they automate the work of probing a network for weaknesses so that those weaknesses can be fixed. An AI-driven version pushes more of that decision-making onto a model rather than a human operator. The same automation that speeds up an authorised test also speeds up an unauthorised one, which is why offensive security tooling has a long history of being repurposed by attackers.

This incident fits that pattern rather than describing a flaw in ARTEX itself. Nothing in the disclosure points to a vulnerability in the tool, and no vendor patch or advisory has been reported. The root cause, as recorded, is misuse of a legitimate tool. That distinction matters for defenders: there is nothing to update, so detection has to focus on how the tool behaves inside a network.

CrowdStrike Intelligence has not published, in the material available, an attribution to a named group or state, and The Agentic Times is not speculating about who was responsible or why South Korean financial firms were chosen. The initial access method is also not described in the available reporting.

The case is an early, concrete example of a trend that security teams have been warning about in the abstract: agentic and AI-assisted security tools lower the effort needed to run a sustained intrusion campaign, and they do so for both sides. An attacker who can point an automated tester at a target network gets reconnaissance, exploitation attempts and lateral movement with far less manual work than before.

Organisations that use AI-driven penetration testing internally may want to treat the tooling as sensitive: restrict who can run it, log every engagement, and make sure monitoring can tell an authorised test apart from an unauthorised one. CrowdStrike Intelligence's full write-up contains further technical detail on the activity.