# South Korean Regulators Confirm AI Agent Breached 16 Banks and Lenders > Source: > Published: 2026-10-17 04:39:13+00:00 *Sixteen South Korean financial firms have been breached since October began. A lone hacker in China is suspected of using an open-source AI pentesting agent to do it, and investigators are still tallying the damage.* South Korea's Financial Services Commission confirmed on October 14 that 16 financial firms had been breached since the start of the month. The list includes Shinhan, KB Kookmin, Hana and BNK Busan banks, along with Yegaram and Welcome savings banks and Hyundai Capital. The regulator put the damage at roughly 100,000 exposed records: names, incomes, loan limits and, in some cases, national ID numbers. Police have assigned 28 cyberterrorism investigators across four teams to figure out exactly how deep it goes. What makes this case different from the usual bank breach isn't the data. It's the tool. Investigators are examining whether the attacks ran through ARTEX, an open-source AI penetration-testing agent that a Chinese security engineer posted to GitHub on July 26, barely two months before the first intrusions surfaced. ARTEX doesn't just scan for open ports. It's built on a multi-agent architecture that identifies targets, finds vulnerabilities, plans an attack path, fires off the actual exploitation tools and checks its own results, largely without a human steering each step. That's according to reporting from BleepingComputer and the cybersecurity outlet The Register. The intrusions didn't hit the vaults. They hit the side doors. Reporting from The Herald Business and SBS News describes attackers going after externally exposed, lower-priority systems: Shinhan's loan-agent inquiry portal, KB Kookmin's mobile employee support tool, Hana's sales support system. These are the peripheral platforms banks build for loan officers and outsourced staff, not the hardened core ledgers. And they're exactly the kind of sprawling, lightly monitored attack surface an automated agent can probe faster than a human red team ever could. At BNK Busan, the exposed data belonged to just 11 outsourced development staff. Even a small vendor contract, it turns out, can become the entry point into a national bank's network. CrowdStrike's investigation adds the most unsettling detail. The security firm traced infrastructure in the campaign to a suspected 26-year-old individual based in Guangdong province, China. It found exposed directories on the attacker's own servers containing ARTEX configuration files alongside session logs from Claude Code, Anthropic's coding tool. According to CrowdStrike's findings, reported by The Register and Infosecurity Magazine, the attacker's setup actually ran DeepSeek's v4.1-flash model as ARTEX's primary engine, with Zhipu AI's GLM-5.3 and xAI's Grok 4.6 layered in for supplementary sessions. The Claude Code logs weren't part of the intrusion tooling at all. They captured the attacker asking Claude where to sell the stolen South Korean data, where to find data-breach Telegram groups, and even to help draft a fake "security researcher" resume, complete with invented credentials and a real location in Maoming. [Developer of hacked bank tool ARTEX pulls it from public GitHub](https://startupfortune.com/developer-of-hacked-bank-tool-artex-pulls-it-from-public-github/) Nine South Korean banks, including Shinhan and KB Kookmin, were hit before ARTEX's creator, known as Autumn-27, pulled the AGPL-licensed tool from GitHub on October 8, citing its misuse in the breach. - [developer pulled ARTEX tool from GitHub after breaches](https://startupfortune.com/developer-of-hacked-bank-tool-artex-pulls-it-from-public-github/) - [open source penetration testing tool causes bank hacks](https://startupfortune.com/developer-of-hacked-bank-tool-artex-pulls-it-from-public-github/) That's a strange kind of carelessness for someone running an AI-automated hacking operation: careful enough to chain four different language models into an attack pipeline, sloppy enough to leave the getaway plan sitting in an open directory. The developer behind ARTEX pulled the project's public GitHub listing and took it closed-source once the Korean breaches were linked back to it, according to Crypto Briefing. That's a predictable reaction, but it doesn't undo the problem. AhnLab, the Korean security firm, has already found ARTEX instances running across roughly 600 IP addresses worldwide, meaning the tool is in circulation well beyond whoever built the original campaign. Pulling one repository down doesn't put that back in the bottle. ## Why regulators can't just patch this away South Korea's Financial Services Commission ordered every bank, card issuer and savings institution in the country to inspect their externally exposed systems on October 2, after the first four breaches went public. That's a reasonable first move, and also an incomplete one. The attacks didn't exploit a single shared software flaw that a patch cycle fixes. They exploited something simpler: large financial institutions run dozens of smaller-tier systems, built by different vendors at different times for different internal teams. An AI agent can now map and probe all of them at a pace no human penetration tester matches. Exact exposure figures still diverge between outlets. Korea JoongAng Daily has reported roughly 66,000 individuals and 2,200 corporate records across seven institutions, while the FSC's October 14 figure spans 16 firms and about 100,000 records. That gap suggests the fuller scope is still being tallied rather than finalized. What isn't in dispute is the shape of the threat. This is the first large-scale, documented case of an AI penetration-testing agent breaching real banks in production, and it took one person in Guangdong, not a nation-state unit, to pull it off. Every CISO who assumed automated intrusion tooling was still a year or two away from matching human red teams just got their answer. **Also read:** [Claude Opus 5.5 beats a human baseline and AI researchers move up their self improvement clock](https://startupfortune.com/claude-opus-55-beats-a-human-baseline-and-ai-researchers-move-up-their-self-improvement-clock/) • [TypeSafe's Jev becomes Vercel AI Gateway's fastest-adopted model ever](https://startupfortune.com/typesafes-jev-becomes-vercel-ai-gateways-fastest-adopted-model-ever/) • [PS5LM Plus now runs AI models natively on a jailbroken PS5's GPU](https://startupfortune.com/ps5lm-plus-now-runs-ai-models-natively-on-a-jailbroken-ps5s-gpu/) *This article is posted in [Technology News](https://startupfortune.com/category/technology/), check it out for more related stories.* ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away.