AlphaGazette
Menu

Security

FBI and AWS warn ARTEX is automating attacks on critical sectors

The FBI said some intrusions gained privileged access in under three hours after targeting began. Amazon honeypots also recorded agents stealing AI credentials and trying deepfake selfie checks at two cryptocurrency exchanges.

3 min read 6 sources AI-generated

The FBI and Amazon Threat Intelligence warn that attackers are using ARTEX, an open-source framework built for AI penetration testing, to automate intrusions against healthcare providers, banks, energy firms, government agencies and other critical sectors. AWS Chief Information Security Officer CJ Moses wrote on Friday that the agents have been taking AI credentials at “industrial scale” to reach models without paying, and making deepfake video in an effort to pass live selfie checks.

Two ARTEX campaigns showed up on sensors in Amazon’s MadPot honeypot from October 3 through November 5. In one, intruders sent AI agents to collect API keys and load them into the framework’s model gateway, which then powered more agents hunting additional keys. Moses wrote that the agents confirmed identities with each provider’s tokens and identity services, listed stored data, judged how far permissions extended, and went after hosted AI model services, stores of secrets and parameters, and application identities. Cloud credentials were used as another route to the models. Named targets included the proxies LiteLLM, Sub2API, Veloera, New API and One API, and the builders Ollama, LobeChat, Open WebUI, RAGFlow, Langflow, FastGPT and Dify.

Deepfakes rejected at two exchanges

A second campaign used ARTEX to produce live deepfake video aimed at know-your-customer selfie checks at two cryptocurrency exchanges, and forged photos of identity documents. Moses said the agents mapped one vendor’s liveness process, recording its face-detection events, the way it follows where a face sits, and the coordinates that direct a person to move. They finished the full movement pattern with the face inside a 7-pixel margin of the aiming point, but the vendor’s server rejected every deepfake. In another session they assembled a live deepfake pipeline on one consumer GPU with an open-source model that animates faces; the clip ran at 11.5 frames per second, a pace Moses called “close to what a live video check expects.” For the document checks, the agents built stand-in camera metadata that copied the real dimensions inside the JPEG and kept the compression tables that came from the camera.

Privileged access and a copy still public

The FBI said ARTEX shortens the time needed to string techniques together, and that some intrusions gained privileged access in under three hours after targeting started. The bureau said the activity can be “noisy,” creating heavy traffic and large amounts of data that may trip defensive alerts, while still moving faster than human defenders can readily match. Attackers often scan quickly for weak spots, then rely on simple flaws such as lax controls or passwords that were exposed or stolen, before spreading toward mail stores, databases and other sensitive records.

The FBI said the tool was used against a U.S. IT giant and South Korea’s financial industry. Its warning also covers a recent supply-chain breach at ConnectWise that left customer environments in Microsoft 365 compromised. The tool has been linked as well to Barracuda and to at least 2,250 ConnectWise customers. Operators can point ARTEX at frontier models hosted by OpenAI or Anthropic, at openly released models, or at models from outside the United States. After it was used in intrusions, the developer closed the source code. The FBI said the older public release is still being passed around and remains “widely accessible.” Researchers last week found hundreds of ARTEX systems exposed on the internet with no password, probably by mistake or misconfiguration, among them the system aimed at South Korea.