How AI-Powered Hacking Outpaced Korean Bank Security

Three Hours. That’s all it took.

Chinese-linked attackers gained administrator-level access to cloud environments at Korean financial institutions using an AI-driven penetration testing tool called ARTEX, according to analysis from Microsoft, Recorded Future, and GsKaler. In several cases, the full privilege escalation from initial compromise to admin control completed before the workday shift changed.

The speed is the alarming part. Traditional supply-chain or credential-stuffing attacks tend to crawl — days or weeks of reconnaissance, failed login bursts, incremental lateral movement. ARTEX flattened that timeline entirely. The tool automates target discovery, brute-force login campaigns, and privilege abuse in one continuous flow. When it hits a misconfigured cloud service, the gap between first contact and full access shrinks to hours.

Who is behind it

Microsoft identified the primary actor as Storm-0940, a Chinese government-linked group active since at least 2021 and previously connected to the Quad7 botnet used in password-spraying campaigns. But Microsoft’s warning was narrower than the picture emerging from other firms.

Recorded Future’s Threat Research Organization Intigriti has tracked a separate entity dubbed TAG-217, which appears to have been testing remote-access infrastructure as early as August 17. GsKaler’s ThreatLabz confirmed ARTEX activity across multiple countries and industries beyond finance, including aerospace, defense, automotive, government, semiconductors, and think tanks — fourteen sectors in total.

That fragmentation of attribution matters. It means multiple groups are sharing, adapting, or independently replicating the same toolkit. ARTEX’s source code was made private after its initial deployment against Korean banks, yet it continues to spread. The tool, written in Go, functions as an AI agent that orchestrates the entire kill chain without human intervention at each step.

How the attack actually unfolded

Microsoft’s analysis of the ConnectWise Cloud backup application breach tells a clear story of tactical evolution. From October 10 to 14, attackers hit 327 accounts across eight organizations from the IP address 144.172.64.99, logging over 151,000 failed login attempts. On October 11 alone, more than 111,000 attempts fired — peaking at roughly 2,900 attempts per minute.

None of those brute-force login attempts succeeded. The attackers hit a wall of strong credentials or lockout policies. So they changed tactics on October 15.

Instead of targeting individual user accounts, they exploited the ConnectWise Cloud application itself. The tool had pre-granted API and service-level permissions designed to let it access customer environments — a legitimate trust relationship turned into a backdoor. By compromising the application’s own authentication credentials, attackers bypassed user login entirely and walked directly into customer cloud infrastructure.

From there, the escalation was rapid. In some cases, attackers modified application authentication configs, manipulated user accounts and groups, and reassigned high-privilege roles — all within three hours of initial access.

Some breaches expanded further, reaching Microsoft Graph and account directories. That gives the attacker visibility not just into a single cloud environment but into the identity layer that governs access across services.

The AI layer most analysts miss

Here’s where the attack diverges from standard nation-state intrusion playbooks. GsKaler’s ThreatLabz found that customized ARTEX deployments were configured to use specific Chinese and Western AI models — DeepSeek, GLM, and Grok — to generate and refine attack prompts.

In some configurations, DeepSeek performed the initial analysis and GLM reviewed or reformulated the output, creating a sort of multi-model attack pipeline. The prompts themselves contained explicit instructions to mass-exfiltrate personal information and transaction data. Records also showed attackers inquiring about the resale value of stolen data.

That last detail is significant. It suggests not every ARTEX operator is a state-sponsored intelligence collector. Some may be financially motivated criminals or hybrid actors blending espionage with monetization. The tool’s availability to multiple groups — not just Storm-0940 — reinforces that possibility.

Why Korean banks were targeted first

South Korea’s financial sector has long been a priority for state-aligned threat groups due to its dense concentration of high-value data, sophisticated digital banking infrastructure, and geographic proximity to China. But the ARTEX campaign represents something qualitatively different from previous operations.

Earlier attacks relied on human-operated credential theft, social engineering, or manual exploitation of known vulnerabilities. ARTEX automates the reconnaissance-to-escalation pipeline and layers AI prompt generation on top. That means the same tool can be tuned for different targets with minimal manual effort, and it can adapt its approach based on real-time feedback from the environment.

The fact that the tool’s developers apparently embedded DeepSeek and GLM directly into the attack workflow indicates a deliberate design choice — not an afterthought. These models can parse security configurations, generate social-engineering prompts, draft exfiltration strategies, and potentially even interpret defensive responses.

What banks got wrong

The ConnectWise case exposes a structural vulnerability that cuts across the entire cloud-services industry. Organizations grant third-party applications broad access to their environments — often with persistent API keys and service accounts that never rotate. When the application vendor is compromised, every customer inherits that exposure.

Microsoft explicitly warned that compromised authentication credentials added by attackers to cloud applications can persist as long-term backdoors, even after the infected software is removed. Rotating credentials alone won’t solve this if the application’s trust relationship with the customer environment remains intact and unmonitored.

Most financial institutions audit user accounts rigorously. Far fewer audit service accounts, application tokens, or the permission scopes granted to third-party vendors. That blind spot is exactly where ARTEX operated.

What happens next

Expect two immediate shifts. First, cloud vendors will face intensified pressure to implement zero-trust architectures for their own integrations — meaning per-customer credential isolation, mandatory rotation, and continuous anomaly detection on service-level access. Second, financial regulators in South Korea and elsewhere will likely expand mandatory cloud-security audits beyond the traditional perimeter model.

The deeper concern is precedent. If a single AI-powered tool can compress a three-month intrusion campaign into three hours across fourteen industries, the economic calculus for both attackers and defenders changes fundamentally. Defenders must now assume that any misconfigured service account or stale API key is a potential admin-level breach waiting to happen — and that the attacker may already be inside before the alert fires.

Korean banks, and the global financial sector more broadly, are now facing a threat landscape where the speed of attack outpaces the speed of detection. The question isn’t whether the next breach will happen. It’s whether anyone will know it happened until it’s too late.