Skip to content- Sources: primary, discussion
- Summary: Amazon Threat Intelligence said on 2026-11-13 that its MadPot decoys, posing as LLM services, received requests from ARTEX agents between 2026-10-03 and 2026-11-05 that carried the operators' instructions and reported findings. The records describe a playbook for stealing provider keys from LLM gateways and app builders such as LiteLLM, One API, New API, Dify, Langflow, RAGFlow, Open WebUI and Ollama, using default admin credentials such as LiteLLM's default master key, unfinished installs, an isCliRequest header bypass in One API family systems, SSRF to cloud metadata, and known Langflow and RAGFlow flaws. Amazon has moderate confidence the operators were separate China-linked groups and cannot verify the outcomes the agents reported.
- Why it matters: Any self-hosted LLM gateway or AI app builder on the internet with default credentials or no auth should be treated as a key store under active, automated attack.
- Background: ARTEX is an open-source agentic penetration testing framework in which a planner agent splits a goal into tasks for worker agents that run security tools and consult an LLM for each next step. LLM gateways such as LiteLLM and One API sit in front of many model providers and hold their API keys, so one weak admin page can expose keys for several accounts. MadPot is Amazon's network of decoy sensors, and here it captured the agents' own context, which includes campaign history, rather than only network traffic. The digest covered the FBI FLASH on ARTEX abuse on 2026-11-15.
- Comments: The Hacker News adds that Cloudflare responders saw ARTEX use a stolen, non-expiring account-level API token to create Worker scripts named artex-*, and that Cisco Talos links more than 2,500 IP addresses to ARTEX.