Urgent Alert

ARTEX: The Open-Source AI Pentest Tool Turned Hacking Weapon

A framework built for authorized penetration testing, now used by real threat actors against real banks. The FBI issued a FLASH advisory on November 13, 2026. Here's what's confirmed, checked against several independent sources.

November 19, 2026 · SunuCyberSecurity Team

Partager :WhatsAppFacebookLinkedInX
< 3h

from initial targeting to confirmed privileged access, per the FBI's FLASH advisory

1,600+

publicly exposed ARTEX consoles found on the internet in early November 2026, some with no password at all

~67,000

people whose data was exposed in the breach of three South Korean banks

200+

organizations targeted across 14 countries in 11 days by a single Chinese espionage group (Vapor Panda)

Sources: FBI FLASH advisory (IC3, November 13, 2026), BleepingComputer, TechCrunch, CrowdStrike — full list at the end of this article

On November 13, 2026, the FBI issued a FLASH advisory about ARTEX, an open-source, AI-driven penetration testing framework. The finding is stark: threat actors, including at least two groups linked to China, are using a tool built for authorized intrusion testing to automate real attacks — against banks, law firms, government agencies, and cloud providers. Three separate, independently documented incidents show what that looks like in practice.

1. ARTEX: a legitimate pentest tool, until it wasn't

ARTEX was published on GitHub in July 2026 by a developer known as Autumn-27, under the AGPL-3.0 open-source license. Written in Go, the tool has no AI model of its own: it connects to third-party models — Anthropic's Claude, OpenAI's GPT, as well as DeepSeek, GLM, Grok, or Kimi — to turn a high-level goal ("test this domain's exposure") into a sequence of concrete actions carried out by specialized agents: reconnaissance, vulnerability discovery, exploitation, privilege escalation, data exfiltration.

On paper, the architecture is designed to stay auditable: a planner breaks the goal into intents, "worker" agents execute them with real tools (Kali, Bash, HTTP requests), and every result is tied to a shared graph of discovered assets plus a per-task exploration graph — with captured HTTP traffic serving as evidence for each finding. Safeguards exist in the code: mandatory human approval for certain commands, scope rules, token and turn budgets. None of that stops a malicious operator from simply disabling those safeguards and pointing the tool at unauthorized targets — exactly what the incidents below confirm.

On October 8, 2026, following abuse reports, the original repository went closed-source. But the open version keeps circulating widely: community forks, archived copies, localized versions (a Korean adaptation exists among others), and even a commercial edition called ARTEX PRO, sold separately. The FBI is explicit on this point: taking down the repository changed nothing about the tool's real-world availability, which remains widely accessible.

2. The FBI's November 13, 2026 advisory

The FBI published a 13-page FLASH advisory ("Threat Overview of ARTEX Autonomous Penetration Testing Framework," reference FLASH-20261113-001, TLP:CLEAR — cleared for public release) via IC3.gov. The advisory documents malicious use of ARTEX since at least August 2026, against a wide range of sectors: financial services, government, technology, defense, healthcare, telecom, law firms.

The central finding: some intrusions went from initial targeting to confirmed privileged access in under three hours. The FBI notes that ARTEX's traffic can be picked up by detection tools — it does trigger alerts — but its speed leaves defenders little time to respond before the attack has already succeeded. Researchers counted more than 1,600 publicly exposed ARTEX consoles on the internet in early November, with the real number of deployments likely much higher — and some of those consoles are protected by no password at all.

3. Three real incidents, three faces of the same threat

South Korea: three banks, around 67,000 people affected. Shinhan Bank, KB Kookmin Bank, and Hana Bank were hit by a Chinese-speaking threat actor combining ARTEX with Claude Code agents. Personal data belonging to roughly 67,000 people was exposed, per the FBI (TechCrunch cites "at least 68,000" — the gap is minor and likely reflects different counting points rather than a real disagreement). CrowdStrike investigators found the attacker's exposed infrastructure, including ARTEX configuration files and Claude Code session histories left openly accessible, which helped reconstruct part of the attack. The root cause cited: weak defenses and no multi-factor authentication on the affected systems. A separate incident in the same country: South Korean police arrested seven suspects linked to attacks on more than 90 organizations, with ARTEX used in most cases — around 8.2 million records stolen from hospitals, schools, and businesses.

ConnectWise supply chain: 2,250+ Microsoft 365 customers targeted. Attackers compromised application identities used by ConnectWise Cloud Backup (formerly SkyKick Cloud Backup), then abused permissions already granted to those applications to authenticate directly into customer tenants. Huntress counted at least 162 compromised tenants. Microsoft attributes the campaign to Storm-0940, a China-aligned actor, and confirms ARTEX was used to attempt authentication against more than 2,250 downstream Microsoft 365 tenants — some attempts progressing to credential changes and persistence attempts.

Vapor Panda: 200+ organizations targeted across 14 countries in 11 days. CrowdStrike documented an espionage campaign cautiously attributed to the Chinese group Vapor Panda, starting October 17, 2026. Over eleven days, more than 200 organizations across 14 countries were targeted by AI agents primarily powered by DeepSeek, supplemented by a second, subscription-based commercial model CrowdStrike could not identify. An important nuance: that 200+ figure describes organizations targeted, not compromised — CrowdStrike confirms actual post-exploitation activity at only nine organizations, including foreign ministries in Latin America, with files already staged for targeting Central America and Asia. CrowdStrike notes this campaign is separate from the Storm-0940 activity seen at ConnectWise.

What we know — and what we don't

The "targeted" vs. "compromised" distinction changes everything. The most dramatic number in this story — 200+ organizations across 14 countries — describes the reach of Vapor Panda's automated sweep, not the number of actual victims (9 confirmed). Both numbers are accurate, but they answer different questions; conflating them overstates the campaign's real scale.

The Vapor Panda attribution stays cautious. CrowdStrike uses hedged language and does not publish a formal confidence level for this specific link — a rigor we'd rather flag than smooth over for a punchier headline.

The tool didn't disappear when the original repository closed. On October 8, 2026, the public source code was taken down — but forks, archived copies, localized versions, and the commercial ARTEX PRO edition keep circulating widely, exactly as a fork we came across ourselves on GitHub confirms.

What to do now

Turn on MFA everywhere, no exceptions

The breached South Korean banks had exactly this gap: weak defenses and no multi-factor authentication on exposed systems — the simplest fix, and the most commonly skipped.

Audit permissions granted to third-party applications

The ConnectWise incident shows that one forgotten application permission, never revoked after a vendor relationship ended, is enough to leave a backdoor exploitable months later.

Watch for bursts of automated requests

The FBI itself describes ARTEX's traffic as detectable — it does trigger alerts — but its speed leaves defenders little room to react. An abnormal spike of authenticated requests deserves immediate investigation, not an end-of-day ticket.

Never expose an admin console with no password

That's exactly the gap that let researchers find more than 1,600 ARTEX consoles themselves, sitting open on the internet.

Don't assume size or location protects you

The targeted sectors are broad — finance, government, healthcare, law firms, telecom — and Vapor Panda hit 14 countries across multiple continents in just 11 days.

Report any suspicious ARTEX-related activity

The FBI encourages reporting through IC3.gov, useful even outside the US to feed shared intelligence on attack infrastructure.

Sources and methodology

FBI / IC3, FLASH advisory FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework" (primary source) · BleepingComputer, summary of the FBI advisory and the ConnectWise incident · TechCrunch, independent cross-check (victim figures, ConnectWise and South Korea incidents) · CrowdStrike, technical analysis of the Vapor Panda campaign. Data as of November 19, 2026 — investigations are ongoing, and victim and organization counts may change; check primary sources for the latest updates.

Related articles

nginx, Apache Roller, OpenJDK: Three Critical Flaws in 48 Hours PHP2Shell: Two Critical RCE Flaws Force an Emergency PHP Patch

Check your own exposure

A free scan, in minutes, no credit card required.

Free scan

Newsletter

Get our next analyses

One email whenever we publish a new article. No spam, one-click unsubscribe.