The U.S. FBI warns that ARTEX, an open-source, AI-driven penetration testing framework, has become a weapon for cybercriminals and nation-state actors alike. Some intrusions have moved from initial targeting to privileged access in less than three hours.
The FBI FLASH released November 13 documents on ARTEX campaigns dating back to at least August 2026 against organizations in financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications, and legal services. It also makes clear the problem isn't winding down.
The original developer pulled the GitHub repository in October, but forks, mirrors, and a commercial version called ARTEX PRO keep the framework in circulation. Researchers counted more than 1,600 publicly-exposed ARTEX consoles in early November. The FBI's indicators also include infrastructure from a newer, financially-motivated campaign that Microsoft tracks as Storm-3358, still active as of November 6.
A developer using the handle "Autumn-27" published ARTEX on GitHub in July as a tool for authorized security testing. Written in Go, it turns natural-language objectives into tasks for a team of AI agents. The agents then work through reconnaissance, exploitation, privilege escalation, and data collection with little human input once the goal is set. ARTEX has no model of its own; it plugs into outside LLM providers. According to the FBI, attacker-controlled instances lean heavily on DeepSeek's Flash models because they are cheap, and showed no observable guardrails when tasks were framed as penetration testing.
Two campaigns put ARTEX on defenders' radar. In early October, CrowdStrike reported that an actor had used it against South Korean financial organizations; the FBI says that campaign exfiltrated personal data on roughly 67,000 people. Two weeks later, Microsoft tied ARTEX to Storm-0940, a Chinese state-aligned group, and to the compromise of ConnectWise Cloud Backup, which gave the attackers access to more than 2,250 downstream Microsoft 365 tenants. Microsoft has also cautioned that many unrelated actors use the framework, so ARTEX activity alone says little about who is behind an attack.
The ConnectWise case shows where ARTEX does its real damage. Between October 10 and 14, Storm-0940 ran a password-spray campaign against 327 accounts at eight organizations, and Microsoft logged more than 151,000 failed sign-ins with no successful logins. The intrusion that worked, beginning October 15, didn't need a password at all. The attackers used compromised credentials for ConnectWise applications that customers had already approved, and they authenticated as those trusted apps without any user signing in.
"ARTEX is not breaking in so much as logging in with keys it finds lying around," said Morey Haber, Chief Security Advisor at BeyondTrust, "including machine identities and service accounts that nobody is monitoring or that have been forgotten about." He frames the threat as an identity and privilege problem rather than a malware problem. Because ARTEX maps trust relationships between organizations, he added, a company is only as secure as its weakest supplier.
That shift changes where defenders need to look. Black Duck's Collin Hogue-Spears, senior director of solution management, notes that Microsoft Entra ID logs application sign-ins separately from user sign-ins, so a team watching only user activity could miss this kind of access entirely. He recommends flagging any vendor application that signs in from outside the vendor's expected infrastructure. The FBI's guidance lands in the same place. It tells organizations to review successful and failed service-principal sign-ins separately, to search for affected apps by application ID rather than display name, and to look beyond the published IP indicators.
ARTEX excels, the FBI says, at finding systems that get less security attention than flagship applications: partner and supplier portals, mobile backends, employee and sales 0upport tools, legacy services, third-party integrations, and apps built by individual business units. The entry points listed in the advisory read like basics gone wrong: missing authentication on web APIs, hardcoded administrator passwords in JavaScript bundles, SQL injection, weak admin credentials, and leaked VPN configurations.
None of those are an unpatched CVE. "Patching closes flaws in software you buy," Hogue-Spears said, "but when a hardcoded password sits in code your own team wrote, no vendor ships a patch for it." His fix is as much organizational as technical. Every internet-facing application, partner portal, and legacy service should have a named service owner and recurring authentication and authorization tests. If nobody can say who owns the partner portal, he argued, its flaws stay open.
ARTEX isn't stealthy. The FBI notes that it generates high volumes of forensically noisy traffic that may set off alerts. The catch is speed, which leaves little time for a coordinated response before the attackers reach their objective.
The advisory's details show both sides. In observed intrusions, ARTEX enumerated more than 110,000 Active Directory objects and checked which Windows machines lacked CrowdStrike Falcon sensors. Its attempts to move laterally to domain controllers, however, were blocked by endpoint protection. Where controls were deployed, they held; the agents went looking for the gaps.
With some intrusions reaching privileged access in less than three hours, how fast can a team act on what it sees? "If your change board has to sign off before anyone can disable a compromised application," Hogue-Spears said, "find out whether that sign-off takes longer." Haber makes the same case for behavioral analytics and automated response, warning that an attack window of a few hours "will not wait for a ticket queue."
Speed shouldn't come at the expense of evidence. The FBI advises keeping affected ConnectWise application connections disabled until ConnectWise issues restoration instructions. It also warns against deleting enterprise applications as a first response, so that configurations and forensic evidence survive.
ARTEX runs on access to someone's AI models, and that leaves a trail. The FBI asks victims to report outbound connections to LLM API endpoints, self-hosted LLM gateways or reseller software, and unusual token consumption or API billing. Check Point research, which examined the databases of 10 exposed ARTEX servers, found that model costs for the campaigns it observed ran from the low hundreds of dollars to about $1,500. Hogue-Spears suggests treating an unexplained jump in AI model spending as a possible sign of stolen API keys.
Drawing on the FBI's recommendations and the experts' guidance, security teams should prioritize the following.
Inventory every internet-facing application. Include partner portals, mobile backends, legacy services, and business-unit apps, and assign each an owner.
Lock down third-party app access. Apply least privilege to multi-tenant app permissions and consent grants. Monitor for credential additions and unusual app-only activity on highly privileged applications.
Eliminate standing privilege. Replace permanent access with just-in-time elevation, so a stolen credential expires before it’s useful.
Hunt for exposed secrets. Scan code repositories, CI/CD pipelines, infrastructure-as-code, and container images for embedded credentials, and rotate anything exposed.
Rate-limit your portals. Alert on sequential lookups and unusually high numbers of records accessed per session on partner and employee portals.
For Diana Kelley, CISO at Noma Security, ARTEX is less a new kind of threat than a bill coming due on old ones. Companies carry years of legacy security debt, she said, and "as those weaknesses become cheaper to find and exploit, that debt is accumulating interest faster."