News · 2026-11-16
FBI warns that ARTEX agent software featured in real intrusions
The FBI warns that attackers have used ARTEX, an AI-assisted agent platform, in intrusions tied to South Korean financial organizations and a cloud backup supply-chain compromise. Microsoft separately reports attempted sign-ins to more than 2,250 customer tenants in the backup-related campaign; that number does not establish how many were compromised.
Key facts
- The FBI issued its public advisory on November 13, 2026.
- Microsoft reports attempted sign-ins to over 2,250 tenants in the ConnectWise-linked activity.
- ARTEX uses external language models and specialized software agents.
- Primary source: the FBI advisory.
Security teams have long faced scripts that can execute thousands of routine actions. ARTEX adds a planning layer: according to the FBI, the Go-based platform translates broad natural-language goals into instructions for specialist agents, which in turn operate tools. Think of an operator assigning jobs to a dispatcher, which hands subtasks to workers. The underlying tools and permissions still determine what actually happens.
The FBI's TLP:CLEAR notice describes activity including account and cloud discovery, credential abuse, web-shell deployment and data collection. It also supplies indicators and defensive guidance; indicators can become stale or be reassigned, so their presence is a starting point for investigation rather than a verdict by itself.
Microsoft's investigation attributes the ConnectWise Cloud Backup compromise and subsequent ARTEX-assisted activity to Storm-0940. Microsoft says the attacker attempted sign-ins to more than 2,250 tenants and reached privileged access in multiple customer environments, in some cases within three hours of initial targeting. The attempted-sign-in figure is a count of targets, not confirmed victims. Microsoft cautions that an affected application or even successful authentication does not by itself prove customer data was accessed or taken.
One separate password-spraying phase generated more than 151,000 failed logon events across 327 accounts at eight organizations. Microsoft says that phase produced no successful authentication or token issuance. The failed phase and the other cases of privileged access are distinct pieces of the investigation, not competing descriptions of the same result.
CrowdStrike's separate account describes ARTEX-related infrastructure and model-session records associated with a campaign against South Korean financial organizations. It reports data exfiltration and says that instance used DeepSeek as its primary model backend. The FBI estimates that information about approximately 67,000 people was exposed; CrowdStrike's public account does not independently confirm that total or name an adversary.
The practical change is that a language model can help an operator coordinate familiar tools at speed and scale. Defenders should examine application permissions, unusual sign-ins and privileged activity alongside the indicators in the advisory. The evidence does not show that a model independently selected every target, exploit or escalation step, nor that the named cases prove a general rate of autonomous compromise. Understanding the agent harness helps separate model behavior from the tools surrounding it.
What defenders can infer
The campaign has two separate lessons for organizations using cloud services. First, a trusted third-party application can become a route into many customer environments if its permissions are abused. A tenant inventory and review of the privileges granted to connected applications help identify where an attacker could move after a supplier is compromised. These are ordinary identity and supply-chain controls, made more urgent when an attacker can automate the work around them.
Second, attempted authentication and confirmed data theft require different evidence. An attempted sign-in can justify investigation; it does not prove a token was issued. A successful login can be an important escalation but does not, on its own, demonstrate that files were read or exported. Investigators need the relevant authentication, application-consent, privilege and data-access records to distinguish the stages.
The two public investigations also differ in attribution. Microsoft names Storm-0940 for the ConnectWise-linked incident. CrowdStrike describes the South Korean campaign without attributing it to that named actor. Combining the accounts into one supposedly single global campaign would therefore go beyond what either company establishes. The FBI advisory gives defenders a common starting point while cautioning that indicators age and can be reused. Those distinctions matter because the safest response is driven by the evidence in each environment, not by the largest number in a headline.
Key questions
Were 2,250 tenants confirmed compromised?
Does ARTEX contain its own AI model?
Were all attack steps autonomous?
Cite this
APA
Ground Truth. (2026, November 16). FBI warns that ARTEX agent software featured in real intrusions. Ground Truth. https://groundtruth.day/news/fbi-artex-ai-agent-attacks.html
BibTeX
@misc{groundtruth:fbi-artex-ai-agent-attacks,
title = {FBI warns that ARTEX agent software featured in real intrusions},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/fbi-artex-ai-agent-attacks.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.