Low-confidence report · Case No. FB-2026-09-27-deepseek-artex-korean-banks
An unnamed, likely Chinese-speaking actor ran ARTEX, an open-source AI pentesting agent with DeepSeek V4.1 Flash as its main model, against South Korean banks' peripheral systems, exposing data on tens of thousands of customers.
This report is hidden because you're viewing high- and medium-confidence reports only. It isn't counted in the scores you're seeing.
Leaderboard / Case No. FB-2026-09-27-deepseek-artex-korean-banks
An unnamed, likely Chinese-speaking actor ran ARTEX, an open-source AI pentesting agent with DeepSeek V4.1 Flash as its main model, against South Korean banks' peripheral systems, exposing data on tens of thousands of customers.
Score breakdown
| Component | Value | Reasoning |
|---|---|---|
| Sentence-Years | 5 | 18 USC 1030(a)(2)(A). Obtaining customer financial data from banks' systems without authorization is 18 USC 1030(a)(2)(A); CrowdStrike assesses the actor as financially motivated and the actor asked about where Korean breach data is sold, supporting the five-year felony maximum for private financial gain under 1030(c)(2)(B)(i). |
| × Contribution | 2 | Ran the operation: CrowdStrike says "the ARTEX instance used DeepSeek v4.1-flash as the primary LLM backend" and that the instance was "likely responsible for the described Korean attacks"; ARTEX is an autonomous penetration-testing agent, so the model ran the intrusions under the human's direction. |
| × Blast Radius | 1 | Third party: The victims are South Korean commercial banks and lenders; the breached systems were peripheral (a loan-progress inquiry service for brokers, an employee mobile work-support system, a sales-support system), not government systems. |
| × Legal status | 1 | Crime. |
| + Tradecraft | 0 | Korea Herald reports the attacker "bypassed identity verification" in Shinhan's loan-inquiry service and Kyunghyang describes presumed automated brute-forcing of customer numbers; neither is a stolen credential, exploit or other rubric technique as reported, and CrowdStrike lists only infrastructure, AI-capability and proxy techniques. |
| + Guardrails | 0 | Guardrails intact: No jailbreak or modification of the model is reported; the actor gave ARTEX a Chinese-language pentesting prompt and reached DeepSeek through a likely API reseller (xcai[.]pro), so guardrails are scored intact. |
| + Dwell Time | 7 | Kyunghyang's timeline has the attack on KB Kookmin beginning on 27 September; Korea Herald says Shinhan confirmed its breach and KB Kookmin detected a possible breach on Wednesday 30 September, so 3 days, detected by the victims. |
| × Recidivism | 1.25 | Same model family offended within 90 days. |
| Total | 21.25 |
Charges the human could face
18 USC 1030(a)(2)(A)Up to 5 yearsAct on Promotion of Information and Communications Network Utilization and Information Protection, Art. 48 (unauthorized intrusion)KR