Key Takeaways

  • ARTEX has enabled some intrusions to progress from initial targeting to privileged access in under three hours.
  • Campaigns have exposed data on approximately 67,000 people and affected more than 2,250 ConnectWise Cloud Backup downstream tenants.
  • The FBI recommends prioritizing peripheral applications, cloud identities, business APIs and supplier connections.

The Federal Bureau of Investigation has warned that malicious actors are using ARTEX, an open-source, large language model-driven penetration testing framework, to automate cyber intrusions across multiple industries. Observed campaigns have targeted financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications and legal organizations since at least August 2026.

The central concern is speed. According to the FBI FLASH, some ARTEX-enabled intrusions moved from initial targeting to privileged access in under three hours. That sharply compresses the period in which security teams can validate an alert, coordinate a response and contain an attacker before sensitive systems are reached.

ARTEX is written in Go and uses multiple specialized agents to convert natural-language objectives into machine-executed actions. It does not contain its own AI model. Instead, it connects to external models including Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi. Analysis of attacker-controlled deployments found DeepSeek Flash family models particularly prevalent, reflecting their combination of capability, cost and limited observable guardrails when requests are framed as penetration testing.

The ARTEX framework is not introducing an entirely unfamiliar catalog of vulnerabilities; rather, its significance comes from chaining established techniques quickly, repeatedly and across numerous targets. The framework has been observed scanning internet-facing assets, testing cloud permissions, exploiting application flaws, deploying web shells, manipulating OAuth credentials, enumerating Active Directory and preparing cloud data for exfiltration.

That automation changes the economics of offensive operations. Smaller groups can run parallel campaigns without staffing each target with an experienced operator. High volumes of failed or duplicated requests may create noise, but that noise does not necessarily buy defenders much time.

The consequences are already measurable. Financially motivated attacks against South Korean financial institutions resulted in the exfiltration of personal information belonging to approximately 67,000 individuals. Separately, a supply chain compromise involving ConnectWise Cloud Backup enabled access to more than 2,250 downstream tenants. Microsoft Threat Intelligence connected ARTEX activity with sign-ins and Microsoft Graph API operations inside customer environments.

Peripheral systems appear especially attractive. Partner portals, mobile backends, employee applications, sales-support services and legacy integrations often receive less scrutiny than flagship customer platforms, yet they hold credentials, expose administrative functions or maintain trusted links into more valuable environments. Automated agents can exhaustively inspect hundreds of these overlooked entry points at machine speed.

The framework also makes supply chain relationships easier to map and exploit. Observed activity included abuse of cross-account cloud trust, third-party credentials and trusted B2B federation. Once attackers compromise a weaker supplier, they can use existing permissions or integrations to move toward upstream customers.

Removing the original repository has not contained the technology. Following public disclosure of malicious usage, the original ARTEX developer removed the GitHub repository in October 2026. However, forks, archived copies, localized variants and a commercial version called ARTEX PRO remain available. Researchers had identified more than 1,600 publicly exposed ARTEX consoles by early November 2026.

For defenders, behavior may offer more durable evidence than individual IP addresses. The FBI highlighted rapid permission checks, sequential portal lookups, unusual service-principal sign-ins, sub-second cloud storage tests, Microsoft Graph reconnaissance, unexpected LLM API traffic and sudden credential additions to OAuth applications. These activities map across multiple MITRE ATT&CK categories, from active scanning and trusted relationships to cloud account manipulation and data staging.

Organizations can respond by inventorying every internet-facing application, restricting administrative and publishing interfaces, reviewing app-only permissions and applying rate limits to high-impact APIs. Teams should also preserve Microsoft Entra, Microsoft Purview, SIEM and Microsoft Graph telemetry while pursuing urgent containment. For ConnectWise-related activity, the FBI advises reviewing records from at least October 1, 2026, including failed authentications and volume anomalies. The practical shift is straightforward: security programs need detection and response processes measured in minutes, because ARTEX has shown that attackers increasingly operate on that clock.