Key facts
- The FBI says multiple threat actors have weaponized ARTEX since at least August 2026 against organizations worldwide, including healthcare.
- The FBI warns the framework lets actors with minimal technical expertise run complex multi-target operations simultaneously, expanding attack scale.
- Storm-0940 used ARTEX for reconnaissance, password spraying, and cloud application permission abuse in the ConnectWise campaign.
- After compromising ConnectWise application identities, attackers used ARTEX to try authenticating against over 2,250 downstream Microsoft 365 tenants.
The bureau says ARTEX has been weaponized by multiple threat actors against healthcare and other organizations worldwide since at least August 2026. The bureau warned that weaponized ARTEX targets ten economic sectors, including healthcare, and called the framework an autonomous AI penetration-testing tool.
The bureau says the framework shortens the time required to chain attack techniques, and some intrusions reach privileged access in under three hours. It also lets actors with minimal technical expertise run complex multi-target operations at once, expanding the scale of attacks.
CrowdStrike linked Vapor Panda, a Chinese cyberespionage group, to automated ARTEX-powered attacks on more than 200 organizations in 14 countries. In that campaign, AI agents deployed web shells and tried different attack paths after initial failures, compromising at least nine organizations.
Microsoft said Storm-0940, a China-state-aligned threat actor, carried out the ConnectWise Cloud Backup supply-chain attack in October. During that campaign, Storm-0940 used ARTEX for reconnaissance and password spraying, along with cloud application permission abuse. After breaking into ConnectWise application identities, attackers used ARTEX to attempt authentication against more than 2,250 downstream Microsoft 365 tenants.
Intrusions in the ConnectWise campaign progressed to credential modifications and attempts to create persistent access. Huntress said it identified at least 162 compromised tenants in that incident.
Storm-0940 targeted more than 300 user accounts at eight critical organizations around Oct. 10. More than 151,000 authentication requests were sent at roughly 2,900 attempts per minute, and the campaign did not trigger automated lockout or alert.
Microsoft said financially motivated Cambodia-based cluster Storm-3358 began using ARTEX in attacks on financial organizations mainly in Southeast Asia. The cluster used ARTEX with conventional automated tools. It scanned internet-connected systems, exploited vulnerabilities, and deployed open-source remote access Trojans. Microsoft’s Storm-3358 blog contains many indicators of compromise and proposes defensive honeypots.
Attacks on South Korean financial institutions exposed personal information belonging to approximately 67,000 people across banks such as Shinhan Bank, KB Kookmin Bank, and Hana Bank. South Korean police arrested seven suspects tied to attacks on more than 90 organizations, and ARTEX was reportedly used in most intrusions. The gang took about 8.2 million data records from organizations including hospitals, educational institutions, and businesses.
A Chinese-speaking threat actor combined ARTEX with Claude Code agents. CrowdStrike found exposed infrastructure holding ARTEX configuration files. It also held Claude Code session histories.
Barracuda disclosed an Artex-powered intrusion between November 1 and November 2. About 10 cloud accounts were touched in roughly 44 hours. The intruders left a backdoor in a Kubernetes cluster. They searched for Ethereum keystore files along with seed phrases and extended private keys.
Intruders used an exposed cloud access key to chart the environment, Barracuda said. They reused secrets, and the company has not assigned blame since many groups use ARTEX. The incident resembled a Railway compromise from October and is part of a wave of intrusions across dozens of AI developer platforms, Barracuda said.
Amazon observed two ARTEX-powered campaigns. One stole credentials to access AI models at scale, and another targeted cryptocurrency exchanges. Amazon said ARTEX attackers relied on long-known techniques carried out at greater speeds. They succeeded against default or hardcoded credentials, keys in code, public artifacts, unauthenticated interfaces, and unpatched software.
GitGuardian detailed 45 agentic attack frameworks. It called them ARTEX siblings, and they are capable of autonomous attacks. Some include built-in command-and-control functionality.
ARTEX is a penetration-testing framework built around multiple AI agents and written in Go. The open-source project was first published on GitHub in July 2026 by developer Autumn-27. The original ARTEX GitHub repository was removed on October 8, 2026.
The framework remains available through community forks and archived copies. Localized versions such as ScopeWeaver also exist, along with a commercial edition called ARTEX PRO. The creator said they had nothing to do with attacks linked to ARTEX. The tool was described as intended only for authorized penetration testing.
ARTEX’s author closed the repository to the public. The author halted public development to block misuse.
ARTEX has no built-in AI model and integrates with external large language models such as Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok, and Kimi. ARTEX is written in Go. It has an embedded Next.js web console that offers an offensive command-and-control platform. Attackers supply goals and tasks.
ARTEX turns them into commands. Specialized AI agents run those commands with little human involvement.
FBI analysis found Chinese models are most prevalent in attacker-controlled ARTEX instances. DeepSeek Flash family models dominating usage is part of that finding. FBI analysis also found DeepSeek Flash family models dominate attacker-controlled ARTEX usage. Their low cost and lack of observable guardrails for penetration testing explain the dominance.
SecurityWeek reported DeepSeek v4.1-flash as the initial model. GLM-5.3 and Grok 4.6 appeared in certain instances.
The FBI says ARTEX is especially effective against overlooked systems such as portals for partners and suppliers, mobile back ends, tools for employee and sales support, older services, and connections to third parties. ARTEX has demonstrated supply chain capability. It changed iOS application distribution infrastructure. Exposed cloud credentials were used, and trust relationships were mapped to pivot into enterprise networks.
More than 1,600 ARTEX consoles were publicly exposed and identified by researchers as of early November. Actual deployments are believed to be significantly higher. SecurityWeek reported on Censys findings of roughly 1,500 hosts with exposed Artex consoles in early November. Dark Reading reported that Censys observed more than 1,400 deployments of internet-exposed Artex consoles as of Nov. 4.
Censys estimated that actual global ARTEX use exceeds observed deployments since internet exposure is not required to run the framework. Unit 42 found 609 internet-facing hosts running ARTEX as of October 2026. It also identified 1,323 hosts running CyberStrikeAI, with 21 hosts running both. Health Exec said the true number of ARTEX deployments is likely much larger once systems behind firewalls or private networks are counted.
Infosecurity Magazine reported on November 13 that the FBI Flash notice was published. Health Exec reported on Oct. 13 that the FBI communicated the warning in a FLASH advisory. The FBI’s 13-page advisory lists indicators of compromise like IP addresses, command-and-control systems, and suspicious network behavior.
The FBI cautions that IP addresses listed in its indicators of compromise may have changed since the attacks occurred. It recommends securing internet-facing applications and restricting cloud administration access. The bureau also recommends reviewing application permissions and authentication logs, along with implementing rate limits on sensitive APIs.
The bureau also recommends monitoring unusual bursts of automated requests. It recommends watching suspicious cloud activity and unexpected credential changes, along with large data transfers. Organizations at risk should shrink their external attack surface and harden identity and access points. They should also improve application protections and keep evidence, the FBI says.
The bureau wants information on ARTEX campaigns. The requested information includes infrastructure artifacts and authentication logs. It covers application and API activity. LLM indicators, cloud and container activity, and exploitation artifacts are also requested.
Organizations that identify suspicious ARTEX activity should report findings to the FBI. They can also contact its Internet Crime Complaint Center.
MalwareMustDie made available an ARTEX detection signature pack. The pack uses open-source rules. The rules let defenders find attacks using the framework. MalwareMustDie made the pack available on Friday.
The pack includes rules and indicators. They cover suspicious network requests and identity abuse. Cloud activity, malicious files, and automated attack patterns are also covered. MalwareMustDie published a detection playbook.
The playbook sorts indicators by confidence level. It warns that IP addresses and infrastructure indicators may become outdated. They may also cause false positives.
Unit 42 says ARTEX is open source and widely forked. That means indicator-based detection cannot keep up. Defenders should build detections around operator behavior, the firm says. Unit 42 observed attackers using ARTEX.
They turned public password reset pages into paths to cloud access. Exposed APIs also became paths to cloud access. This sometimes happens in less than a day. Many organizations and sectors saw this activity.
None of it exploited a single software vulnerability.
An ARTEX operator can test every exposed workflow at once, Unit 42 said. The operator can also respond within one minute of finding a stolen credential. Unit 42’s Incident Response team investigated intrusions where the attacker’s server hosted an ARTEX console. It also investigated intrusions where the server used ARTEX infrastructure along with user agents.
Unit 42 Managed Threat Hunting followed a coordinated campaign using ARTEX. Customer cloud environments around the world were targeted, including retail and consulting sectors. Any internet-exposed asset can be hit by automated persistent intrusion attempts. The organization’s size or sector does not matter.
Current autonomous frameworks favor fast trial-and-error over stealth, Unit 42 said. Cairn produced hundreds of thousands of noisy requests from one IP in a study. Defenders can make AI agents waste resources. They can serve identical generic error codes.
They can enforce silent connection throttling. Both approaches deprive agents of contextual cues.
In one observed engagement, the attacker abandoned scripted clients for a desktop browser. This happened during a password reset and first token request. It pointed to human involvement. Financially motivated actors can launch and manage dozens of complex concurrent intrusions for minimal compute cost.
They use off-the-shelf tooling combined with LLMs such as DeepSeek.
Fourteen penetration-testing and red-teaming tools were analyzed by Unit 42. Twelve emerging tools were highlighted alongside CyberStrikeAI and ARTEX across four categories of AI-assisted activity. Unit 42 recommends removing public attack maps like OpenAPI specifications and source maps. It also recommends requiring authentication on every write endpoint, protecting federated identities, and moving function secrets to a secrets manager.
Unit 42 advises hunting for canary usernames in identity-provider logs. It also advises watching for a first cloud key from a new IP and write-then-delete storage tests. ARTEX campaigns show acceleration in speed, coordination, and autonomy, Steve Winterfeld of Akamai said. Activities can be chained with limited hacking skills, he said.
“What is particularly noteworthy about the Artex campaigns is the acceleration in speed, coordination, and autonomy,” he says. Daniel Shechter of Miggo Security said the attacker sets the pace whenever every defensive step must wait for a person. “When every defensive action has to wait for a person, the attacker controls the pace.”

Join the discussion