Some operations using Artex, an artificial intelligence-based security testing tool, went from selecting an organization as a target to gaining high-level access to its systems in under three hours, according to an FBI alert. The agency says attackers have used it since at least August 2026 against banks, government agencies, technology companies, and healthcare providers.
The FBI notice was published on November 13 and circulated while a Microsoft 365 outage and a nationwide flight stoppage by the U.S. Federal Aviation Administration were drawing attention. An FBI official told ABC News that Artex did not appear to be connected to those incidents or to other recent outages, including those affecting banks such as JPMorgan Chase. The FAA said it had found no evidence of a cyberattack.
On November 17, the FBI clarified that the notice was not new and that it had not issued another warning since its publication. The agency said the recommendations in the alert remained in effect. The timing of the outages therefore does not amount to attribution: the official denied that Artex appeared to be behind those events.
What Artex does during an intrusion
Artex automates security testing that organizations use to find weaknesses. According to Microsoft’s analysis cited by the FBI, its agents can assess defenses, exploit flaws, and extract data. Automation makes it possible to test several routes into a network at once, reducing the time defense teams have to respond.
TechCrunch describes a workflow in which Artex quickly scans a network, looks for weak protections or exposed or stolen credentials, and tries to establish initial access. It can then help the attacker move through the network in search of databases, mailboxes, and other sensitive information. The FBI warned that these operations can generate a lot of traffic and trigger alerts, but their speed can leave defenders reacting after the intrusion has already advanced.
Google Threat Intelligence Group (GTIG) and Mandiant detected the use of Artex and other agent frameworks in incidents investigated since late August 2026. Their analyses describe automated tasks at different stages: reconnaissance, lateral movement, database exploitation, and preparation for data exfiltration. Individual Artex sequences could run in bursts lasting between 2 and 25 seconds, according to the activity log reviewed by Google.
Defenders who rely on spotting malicious activity and then responding to it will often arrive after the attacker has finished.
— Arve Kjoelen, Barracuda’s chief information security officer
ConnectWise, Microsoft 365, and victims in South Korea
One of the largest cases involves ConnectWise, a U.S. information technology management company. Microsoft warned that attackers used stolen credentials from ConnectWise backups together with Artex to access Microsoft 365 environments at more than 2,250 organizations. The FBI confirmed the activity. Microsoft said the attackers were likely Chinese, while ConnectWise told TechCrunch it could not confirm the number of affected environments cited by the FBI.
The access relied on credentials for backup software that customers had previously authorized. This allowed intruders to take advantage of existing permissions without signing in as an individual user. Microsoft noted that a successful sign-in alone does not prove that data was accessed or stolen.
The FBI also cited attacks against South Korean financial institutions. CNET puts the number of people affected by the exposure of personal information at approximately 67,000; TechCrunch reports at least 68,000. Both publications describe the theft of personal data, but give different figures, so their accounts do not provide a single definitive number.
TechCrunch also reported an attack against Barracuda attributed to financially motivated criminals. Artex found an exposed credential for a third-party service that Barracuda had not revoked when its relationship with that provider ended. That credential enabled access to key stores Barracuda used to collect data from its customers’ systems.
Barracuda could not rule out access because its logs did not provide enough detail and decided to “assume the worst”: it changed all its customers’ credentials. The company said it notified a minority of its XDR customers and recommended that they revoke and recreate integration credentials, without specifying how many.
An ecosystem of open-source and derivative tools
Artex is not the only framework of this kind. In its November 16 update, Google also listed CAIRN, CyberStrikeAI, and PentAGI, along with HexStrike AI and custom-developed tools. GTIG observed CAIRN and PentAGI in real-world operations, as well as experiments with CyberStrikeAI and HexStrike AI. According to the company, these systems combine specialized agents, language models, and security tools to automate different stages of a campaign.
Google describes Artex as a penetration testing and command-and-control platform written in Go, with a Next.js web console. The framework distributes work among six specialized subagents and can connect to commercial and open-source models. Observed capabilities include Active Directory reconnaissance, privilege escalation, searching for Azure storage keys, and creating and installing custom webshells.
The creator originally published the project on GitHub and took down the repository in October 2026 after attackers abused it. However, the FBI warns that copies continue to circulate and are widely accessible. Google also mentions community variants such as ScopeWeaver and a commercial version called ARTEX PRO.
The tool appeared publicly in July, according to TechCrunch, and its use in attacks has been observed since at least August, according to the FBI and Google. TechCrunch also reported that researchers had found hundreds of Artex systems exposed to the internet without a password, likely due to configuration errors or accidental deployments. One of them was the system used to attack South Korean financial institutions.
Google sees faster attacks, but also more visible ones
Google researchers describe operations in which Artex received compromised credentials and executed later stages of the intrusion. In one case, the framework generated 27 events in 23 seconds while testing AWS permissions across different security groups. In another, it adapted to a script-size error by trimming the content and launching it again; this deployed an instance with 17 snapshots in approximately 10 seconds.
Speed does not always translate into stealth. Google found trial-and-error executions and a volume of activity that can leave useful traces for investigations. In one incident, access to an exposed Linux web server and the execution of SQL queries through API calls generated 240 log events. Researchers observed repeated attempts to create webshells with different filenames until some returned valid results.
Google warns that this footprint can make detection easier, although it does not eliminate the frameworks’ ability to achieve their objectives. The company believes that as the tools and models improve, human intervention could decrease and stealth could increase. In its quarterly analysis, it also notes that attackers are moving from manual instructions to autonomous workflows involving multiple agents.
In a separate campaign, Microsoft warned of access to more than 2,250 organizations through ConnectWise. Google, meanwhile, describes the use of automated tools to find and exploit vulnerabilities. Its report also mentions the rapid exploitation of already-patched flaws and an increase in automated research processes, although these cases should not be confused with a single Artex campaign.
FBI recommendations and Google’s defenses
The FBI recommends limiting access to sensitive systems, reviewing software permissions, replacing exposed credentials, and monitoring unusual account activity or large data transfers. These measures are aimed at organizations that administer the affected systems and accounts; they do not indicate that all Microsoft 365 users have suffered an intrusion.
Google says its security operations team deployed behavioral and volume-based rules to detect the use of Artex. Mandiant published a hardening guide and a defensive architecture framework designed to counter the speed and scale of these tools. In an earlier post, Google listed measures such as isolating the control plane, using short-lived credentials, applying least privilege, segmenting systems, and detecting anomalous behavior.
The FBI alert, published on November 13, remains the current public guidance, according to the agency itself. Meanwhile, the original Artex repository is no longer available, but copies and variants continue to circulate; the alert does not announce the complete removal of these versions.
Sources: CNET, TecnoAndroid, Google Cloud Blog, Google Cloud Blog, TechCrunch