Cybersecurity researchers have disclosed details of a targeted campaign aimed at South Korean financial organizations that used an artificial intelligence (AI) pen testing tool named ARTEX to carry out the attacks.
The activity, per CrowdStrike Intelligence, was active from late September to early October 2026, and resulted in data exfiltration from various South Korea-based financial firms, including Shinhan Bank and Yegaram Savings Bank.
“In this activity, the threat actor leveraged ARTEX, a recently released open-source agentic penetration testing (pentesting) tool developed in China, alongside large language models (LLMs),” the cybersecurity company said.
CrowdStrike said it discovered the campaign after it identified a set of open directories hosted at a Hong Kong-based IP address, exposing Claude Code session histories, Claude memory files, and ARTEX configuration files.
The campaign has not been attributed to any known threat actor or group. But evidence points to a suspected Chinese-speaking operator driven by financial gain.
ARTEX is a large language model (LLM) multi-agent-driven autonomous penetration system developed by Autumn-27. Analysis of the Claude Code sessions from the Hong Kong IP address has revealed a two-server architecture –
- The Hong Kong-based IP address functions as the backbone of the campaign
- The IP address “38.244.50[.]120” hosts the ARTEX instance suspected to be behind the attacks on Korean attacks
The ARTEX instance has been found to use DeepSeek v4.1-flash as the main LLM backend, while using Z.ai’s GLM-5.3 and SpaceXAI’s Grok 4.6 to supplement the model. It’s suspected that the threat actor accessed DeepSeek likely via the LLM API reseller “xcai[.]pro.”
“In addition to conducting ARTEX-related operations, the threat actor asked Claude where threat actors typically sell Korean data breach information and asked Claude for assistance in finding Korean Telegram data sales groups,” CrowdStrike said.
In one Claude Code session, the threat actor is said to have fed a prompt that referenced a Telegram account named “@YY520CN” and the name “YY.” Other sessions related to vulnerability research on a Telegram-based NFT gift marketplace have also used the same Telegram username.
“While the personal details included in the prompt likely belong to the threat actor who conducted the ARTEX-related activity, currently available information cannot definitively associate these details with the threat actor,” CrowdStrike concluded.
The breaches have triggered a warning from South Korea’s Financial Services Commission and Financial Supervisory Service, urging financial consumers to exercise vigilance and be on the lookout for potential phishing attacks and loan scams stemming from leaked data.
In an October 8, 2026, message posted on their Telegram channel, @YY520CN said, “I didn’t do anything. South Korea just wants to blame all the untraceable things on me and other innocent ordinary people, so that they can solve all the problems without taking their own responsibility.”
ARTEX Goes Closed Source
The misuse of ARTEX has prompted Autumn-27 to take it closed source, with the developer emphasizing in a statement that the malicious attacks had nothing to do with them. They also said the malicious use of the tool violates the original purpose of the tool.
“ARTEX was originally designed for the purpose of learning and research,” Autumn-27 said. “It aims to help enterprises and organizations conduct security risk tests within the scope of authorized assets and improve security protection capabilities.”
“In view of the reality of tool abuse, the ARTEX project will no longer be updated and will be converted to a closed source. There will be no release of any version or maintenance support in the future.”
SCARLET LOOP Uses AI for Account Takeover
The development comes as ZenoX disclosed details of an agentic credential stuffing and account takeover platform orchestrated by a financially motivated, Portuguese-speaking actor dubbed SCARLET LOOP.
The operation automates the entire process, right from victim selection to executing the login using stolen credentials obtained from infostealer logs and data leaks to hijack accounts at scale. The entire process unfolds in four steps –
- Target discovery and qualification, which involves finding high-value targets using an AI classifier. Primary targets include Brazilian loyalty, corporate incentives, and gift cards platforms.
- Obtaining credentials specific to each target
- Login execution, which employs an AI agent with an anti-detection browser to bypass defenses that detect automation
- Exfiltrating successful credentials in the format “✅ {url} {user}:{password}” to a private Telegram channel
While OpenAI’s GPT-5.6 is used for dork generation (aka search queries) to find authentication pages of companies in a specific sector, GPT-5.5 is used for target classification and rating. The browser agent utilizes DeepSeek-V4-Pro and DeepSeek-V4-Flash.
Also configured are Anthropic’s Claude Opus 4.6, Google Gemini 2.5 Flash, and GLM-5.1 models, with Google’s gemma-4-26B-A4B set up as a local model served at “127.0.0[.]1:1237.”
“An LLM agent with 46 automation tools drives an instrumented Firefox that spoofs canvas, WebGL, time zone, language, geolocation and viewport, with outsourced captcha solving. The agent finds the login page, fills in the form and classifies the result,” the Brazilian cybersecurity company said in a report shared with The Hacker News.
“In the discovery and login phases the agent acts without step-by-step supervision and autonomously maps unfamiliar login surfaces, discovers security weaknesses, and generates purpose-built automation to exploit them.”
Interestingly, the platform also supports an “AUTO Mode” during the login execution phase that removes the AI model from the loop after a set number of successful logins on the same target domain. The idea is to save on tokens and use them sparingly only when an unfamiliar login form is encountered.
“Model reasoning is expensive and slow,” ZenoX said. “It is indispensable on the first visit to an unknown form and unnecessary on the thousandth. The platform was built to pay for intelligence exactly once per target, and then stop paying.”
Analysis of the internet-exposed server hosting the platform has revealed that 12,277,358 credentials were tested, out of which 11,832 credentials have been classified as valid across 3,968 domains.
The findings once again illustrate how threat actors are increasingly adopting AI in their attacks to enhance their operational tempo and capabilities.
