The FBI issued a FLASH on 13 November describing ARTEX, an open-source penetration testing platform driven by large language models, as a tool that a range of malicious actors have turned into an intrusion engine. The bureau says the framework has been used since at least August 2026 against organisations worldwide, and that in some cases an operator went from first targeting to privileged access in less than three hours. The warning matters because the tool compresses work that once took a skilled team days into a largely unattended run, and it remains easy to obtain.
What the FBI says ARTEX is
ARTEX is written in Go and splits an operation across several cooperating AI agents. An operator describes a goal in plain language, and the agents translate it into commands and carry them out, with little further input from a human. The framework ships without a model of its own; it connects to whichever external LLM service the operator chooses, from commercial frontier models to open-weight ones.
According to the FLASH, analysis of infrastructure controlled by attackers found Chinese-developed models most common, led by DeepSeek’s Flash family. The FBI attributes that preference to price, capability and the absence of visible refusals when a request is framed as penetration testing.
The project appeared on GitHub in July 2026 and gathered an estimated 1,500 to 2,000 stars in about ten weeks. Its author said it was meant for authorised, scoped testing and took the repository offline on 8 October after reports of abuse. The FBI says that changed little: forks, localised variants, archived copies and a paid edition called ARTEX PRO keep it in circulation. As of early November, more than 1,600 ARTEX consoles were visible on the public internet, and the bureau expects the real number of deployments to be considerably larger.
Campaigns cited in the FLASH
The FBI lists a wide spread of targeted sectors: financial services, government, technology, defence, semiconductors, energy, healthcare, manufacturing, telecommunications and legal services. Two operations are singled out.
- South Korean finance. A financially motivated campaign against banks and other financial institutions in South Korea led to theft of personal information belonging to roughly 67,000 people. The FLASH ties a proxy network and an LLM API reseller domain to this activity.
- A backup-service supply chain. Attackers compromised a cloud backup product integrated with Microsoft 365, which gave them a route into more than 2,250 downstream customer tenants. The FBI asks affected customers to review activity from at least 1 October 2026.
The FLASH also lists infrastructure for a cluster it labels Storm-3358, including servers for remote access trojans and ARTEX itself that were still active on 6 November. The document does not attribute the overall ARTEX activity to a single group; it describes a variety of actors using the same tool. No criminal charges are mentioned.
How the intrusions unfold
The FBI maps observed behaviour across the full ATT&CK lifecycle, and the pattern is breadth rather than novelty. The agents test stolen cloud credentials for what they can do, probe web applications for common flaws such as injection, server-side request forgery and missing authorisation checks, and hunt for secrets left in front-end code or configuration. Once inside, they plant web shells, add credentials to existing OAuth applications, spin up cloud compute with powerful roles, and harvest Active Directory material before attempting lateral movement. Data is staged in large new volumes and moved out over HTTPS.
Three points stand out for defenders:
- Peripheral systems get hit first. The FBI says ARTEX is especially effective against supplier portals, mobile backends, legacy services and apps built by individual business units: the estate that rarely gets the same review as the main product.
- Trust relationships are a favoured path. Abuse of partner tenants, federated B2B access and vendor credentials appears repeatedly, which maps to Supply Chain Compromise and Trusted Relationship in ATT&CK.
- Noise does not equal safety. The activity is loud and should trigger alerts, but the speed leaves little time to coordinate a response before the objective is reached.
The FLASH includes IP addresses and a domain, which the FBI asks recipients to vet before blocking because cloud addresses get reassigned. Readers should take the current list from the FBI document rather than any copy.
What is still unclear
The FBI does not give a total count of victims, does not say how many of the exposed consoles belong to criminals rather than legitimate testers, and does not describe how operators obtain the stolen cloud and VPN credentials that start many of these intrusions. It also does not say whether ARTEX PRO adds capabilities beyond the open version.
Where to start
- Find your forgotten edge. Build a complete list of internet-facing applications, including partner, supplier and employee portals, mobile APIs and older services, and put management and publishing interfaces behind private access paths.
- Audit app identities in Microsoft Entra ID. Look for new credentials added to service principals, new owners, consent grants and privileged role assignments. Review successful and failed service-principal sign-ins separately, and do not stop at the published IPs.
- Hunt for secrets. Scan repositories, pipelines, container images, environment variables and front-end bundles for embedded keys, then rotate anything exposed.
- Rate-limit business APIs. Alert on sequential record lookups and unusually large result sets on portals, and enforce authorisation on the server for every customer record.
- Watch LLM egress. Unexpected outbound calls to LLM APIs, unfamiliar model resellers or self-hosted LLM gateways on servers that have no reason to use them are worth investigating alongside other signals.
- Preserve before you purge. Keep sign-in, audit and SIEM data, and record configurations before removing rogue credentials or apps, so the investigation survives the clean-up.
Incidents involving ARTEX or similar agentic frameworks can be reported to a local FBI field office or to IC3. The bureau asks for console screenshots, configuration files, authentication logs and evidence of LLM traffic where available.
Frequently Asked Questions
What is the ARTEX framework? ARTEX is an open-source, AI-driven penetration testing platform in which several agents turn a plain-language goal into commands and run them with little human input. The FBI says malicious actors have used it since at least August 2026 to automate intrusions.
Is ARTEX still available after its GitHub repository was removed? Yes. The FBI says forks, archived copies, local variants and a commercial edition called ARTEX PRO keep it in circulation, and more than 1,600 consoles were exposed online in early November.
How can organisations defend against ARTEX-driven attacks? The FBI’s priorities are reducing internet exposure, especially peripheral and partner-facing systems, tightening and monitoring cloud application identities, removing embedded secrets, and rate-limiting high-value APIs. It also asks organisations to preserve logs before making irreversible changes.
Source: FBI Internet Crime Complaint Center (IC3).