Threat actors are actively leveraging an open-source AI penetration testing framework known as ARTEX to automate cyberattacks against banks, healthcare organizations, government agencies, and energy providers, according to coordinated warnings issued by the FBI and Amazon Threat Intelligence.
As reported by The Register, the tool enables threat actors to coordinate complex, multi-stage intrusions at high speed, chaining exploitation techniques together with minimal human intervention.
Industrial-Scale Model Theft and Deepfake Fraud
Between October 3 and November 5, Amazon's MadPot honeypot network detected two distinct operations powered by the framework. In one campaign, adversaries used autonomous agents to harvest application programming interface (API) keys and cloud credentials at an industrial scale. The stolen keys were subsequently fed back into the ARTEX model gateway, allowing operators to deploy additional automated agents to search for further access secrets.
According to AWS Chief Information Security Officer CJ Moses, the automated agents systematically evaluated identities via provider token and identity services, enumerated data storage, evaluated permissions, and targeted managed AI model services and application identities. Targeted technologies included large language model gateways and proxy platforms such as One API, New API, Veloera, Sub2API, and LiteLLM, as well as AI development platforms including Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama.
In a second observed operation, attackers deployed ARTEX to target identity verification checks at two cryptocurrency exchanges. Agents attempted to bypass know-your-customer (KYC) controls by generating synthetic ID documents and real-time deepfake videos designed to beat live selfie requirements.
How the Autonomous Attacks Work
During the cryptocurrency exchange campaign, ARTEX agents reverse-engineered a vendor's liveness detection protocol. The automated system mapped out face-detection event streams, the coordinate system governing movement prompts, and position-tracking algorithms.
Although the agents successfully matched the requested movement sequences within 7 pixels of the target path, the vendor's server-side verification systems rejected the deepfakes on every attempt. In another session, the framework established a real-time deepfake pipeline operating at 11.5 frames per second on a single consumer-grade GPU, running an open-source face animation model.
For fraudulent document verification, agents generated synthetic camera metadata that aligned with the underlying JPEG dimensions and preserved camera compression tables typically lost during manual image editing. The framework coordinated these actions concurrently, assigning separate tasks to different agents, such as clearing balances, preparing deepfake pipelines, and executing verification bypasses.
High-Speed Intrusions and Downstream Impact
Federal authorities warned that the automation significantly compresses intrusion timelines. In a separate alert, the FBI stated that ARTEX allows threat actors to progress from initial scanning and targeting to full privileged access in under three hours.
The FBI connected the framework to an August intrusion against South Korean financial institutions that resulted in the theft of personal records belonging to roughly 67,000 individuals. Additionally, the bureau linked ARTEX to a September supply-chain breach affecting ConnectWise Cloud Backup customers, which led to the compromise of more than 2,250 downstream Microsoft tenant environments.
While the original developer removed the primary ARTEX GitHub repository in October following these breaches, forked repositories, localized builds, and commercial variants remain available. As of November 4, internet scanning platform Censys logged 1,465 exposed ARTEX consoles online, with the FBI assessing that total active installations are substantially higher across private networks and firewalled environments.
What You Should Do
Security investigators emphasized that while the automation expands the speed and persistence of attackers, the underlying techniques rely on exploiting common configuration flaws and poor security hygiene. Defenders should take several steps to harden their environments:
- Eliminate exposed management interfaces: Restrict administrative portals, API gateways, and LLM inference endpoints to internal networks, dedicated administrative pathways, or private connections.
- Implement Zero-Trust controls: Place enterprise resources, model hosts, and self-service portals behind strict identity-aware proxies and zero-trust authentication policies.
- Audit credential storage: Scan public artifacts, codebases, and configuration files for hardcoded API keys, tokens, and parameter store secrets.
- Enforce default credential rotations: Ensure all connected AI builders and open-source proxy deployments are configured without default credentials and have unauthenticated endpoints disabled.
- Apply vendor security updates: Routinely patch exposed server software and containerized platforms to prevent automated vulnerability chaining.





