CVE Tools
A dim data-center server rack with a single monitor glowing, showing a stylized exposed file-directory tree on screen, no logos or on-screen brand text, a faint city skyline visible through a window b

CrowdStrike Found a Hacker's Claude Code Memory File. Five Weeks Later, Korea's Breach Count Hit 37 Companies.

An open-source AI pentest agent called ARTEX, forensic artifacts a sloppy operator left exposed, and an official breach tally that's grown three times since October — while South Korea's own financial regulator still won't say AI caused any of it.

On October 7, 2026, CrowdStrike published research on a financially motivated operator who breached South Korean financial institutions using ARTEX — an open-source, Chinese-built agentic pentesting framework — wired to commercial LLMs including Anthropic's Claude Code. The operator left an open directory on the internet. Inside it, CrowdStrike found the tool's own memory file. That single operational-security failure is the reason any of this is public at all.

Moderate confidenceCrowdStrike's attributionChinese-speaking, financially motivated, no named group
16 → 37financial companies, per South Korea's FSCOct 14 update vs. Nov 5 update
~100K → ~473Kpersonal records exposed, per the FSCsame two updates, three weeks apart
1,600+exposed ARTEX consoles foundFBI FLASH-20261113-001, Nov 13, 2026

What CrowdStrike actually found in the operator's own directory

CrowdStrike describes a "two-server architecture." An IP address, 38.244.50[.]120, hosted the ARTEX instance tied to the Korean intrusions and had an open directory sitting on it. Inside was a Claude Code markdown file — /.claude/CLAUDE.md — holding a Chinese-language prompt instructing the model's conduct, plus a second IP address in Hong Kong. Pivoting to that address turned up more open directories: Claude Code session histories, ARTEX configuration files, and Claude memory files documenting the targeting of specific South Korean financial-sector systems, including, per CrowdStrike, a loan progress inquiry service used by financial brokers at one organization and an employee mobile work-support system at another. CrowdStrike names neither bank; it says the total number of affected organizations "remains unconfirmed" from its own vantage point.

The forensic trail, in the order CrowdStrike reconstructed it

  1. Open directory found — IP 38.244.50.120 hosting an ARTEX instance, left publicly browsable
  2. CLAUDE.md memory file — Chinese-language pentesting prompt + a second IP in Hong Kong
  3. Pivot to Hong Kong server — Primary attacker infrastructure, per CrowdStrike's two-server model
  4. More open directories — Claude Code session histories, ARTEX configs, Claude memory files
  5. Named business systems targeted — A loan-inquiry service and an employee work-support system, at two separate organizations
  6. Operator asks Claude where to sell the data — Queries about Korean Telegram data-sales groups, per CrowdStrike's session review
  7. Operator drafts a résumé citing the operation — Mid-session, unprompted by any defender — the detail that gave CrowdStrike a possible identity lead

Officially, this is still "cannot rule out" — not "confirmed"

Most of the coverage of this story reads as settled fact: an AI agent hacked Korean banks. South Korea's own Financial Services Commission (FSC) has not said that. Its October 14 press release — after 16 financial companies and roughly 100,000 personal records had been confirmed exposed — states plainly that the Financial Supervisory Service and the Financial Security Institute were still examining attack patterns to determine "whether artificial intelligence (AI) has been used in these hacking incidents." Three weeks later, on November 5, with the count up to 37 companies and about 473,000 records, the FSC's language had not moved to confirmation — it frames AI only as a general, forward-looking threat category its new policies are built to address, not a settled cause of the incidents already investigated.

Headline framing vs. regulator framing

What most coverage says
  • "AI hacked Korean banks"
  • ARTEX + Claude Code as the confirmed attack mechanism
  • A single campaign, roughly seven institutions
The framing that traveled fastest.
What South Korea's FSC has actually said
  • AI involvement is "under investigation," not confirmed, as of Oct 14
  • By Nov 5: 37 companies, 40 incidents, ~473,000 records — and AI framed as a trend to defend against, not a confirmed cause
  • No account passwords or payment credentials reported stolen in any update
Two official releases, three weeks apart, neither one closes the question.
MetricFSC update, Oct 14, 2026FSC update, Nov 5, 2026
Financial companies affected1637
Hacking incidentsnot broken out40
Personal records exposed~100,000~473,000
AI confirmed as the cause?No — under investigationNo — framed as a general trend

How a one-operator forensic story became a regulatory mega-incident

  1. ARTEX activity begins, per the FBI
    The FBI's November FLASH dates weaponized ARTEX use to "at least August 2026" — before the Korean campaign started.
  2. South Korean financial institutions breached
    CrowdStrike's own window for the activity it investigated.
  3. CrowdStrike publishes its research
    The open-directory discovery, the CLAUDE.md file, moderate-confidence attribution, no named group.
  4. ARTEX's developer, "Autumn-27," takes the project closed-source
    Stated purpose was research; the developer disclaimed responsibility for how the tool was used.
  5. FSC: 16 companies, ~100,000 records
    AI involvement explicitly described as still under investigation.
  6. Microsoft ties ARTEX to a separate breach
    Storm-0940's ConnectWise Cloud Backup credential theft, reaching 2,250+ Microsoft 365 tenants — a different operator, same open-source tool.
  7. FSC: 37 companies, 40 incidents, ~473,000 records
    AI still framed as a general risk category, not a confirmed mechanism in these specific incidents.
  8. FBI publishes FLASH-20261113-001
    Over 1,600 exposed ARTEX consoles found; the closed-sourcing in October did not stop the tool's spread — forks, archived copies and a commercial "ARTEX PRO" variant remain in circulation.

A takedown that didn't take anything down

ARTEX's own developer closed the source two days after CrowdStrike's report went public. Five weeks later, the FBI's FLASH report treats that gesture as having accomplished essentially nothing: archived copies, independent forks, and a commercial "ARTEX PRO" offering keep the framework available, and researchers had already found more than 1,600 ARTEX consoles exposed on the open internet by the time the FBI wrote its advisory. The same tool shows up tied to a second, unrelated incident we've already covered on this blog — Microsoft's attribution of the ConnectWise Cloud Backup credential theft to the China state-aligned group Storm-0940, which used ARTEX to automate password-spraying and OAuth abuse across more than 2,250 Microsoft 365 tenants. Different operator, different motive, same off-the-shelf tool. That's the actual shift worth tracking: not one dramatic hack, but a framework that keeps resurfacing under different hands faster than anyone can take it down.

What's actually been published as an indicator2 rules
other ×2

No Sigma, Suricata or YARA rule content has been released for this specific campaign. These are the two indicators CrowdStrike and the FBI have put their names to — treat them as hunt leads, not deployable detections.

ARTEX host and open directory: 38.244.50.120GitHub
CrowdStrike Counter Adversary Operations
LLM API reseller domain tied to the campaign: xcai.pro, plus the FBI's full IOC/ATT&CK mapping tableGitHub
FBI IC3 FLASH-20261113-001

Sources: CrowdStrike — Unknown Threat Actor Uses ARTEX to Target South Korean Finance, FBI/IC3 FLASH-20261113-001 — Threat Overview of ARTEX Autonomous Penetration Testing Framework, South Korea FSC — Authorities Hold Meeting on Cybersecurity Management in Financial Sector, Oct 14, 2026, South Korea FSC — FSC Holds Meeting on Strengthening Cyber Defense Framework amid Growing AI Hacking Threats, Nov 5, 2026.