Amazon Threat Intelligence Catches Self-Funding ARTEX AI Agents Stealing LLM Keys
Amazon MadPot decoys captured operating memory from ARTEX AI agents hijacking exposed LiteLLM, Dify, and One API servers to fund further attacks.

By Elena Rostova, Cybersecurity
, updated 3 min read
Threat actors have weaponised an open-source autonomous penetration-testing framework called ARTEX to run self-funding cybercrime campaigns, according to telemetry published by Amazon Threat Intelligence.
Between October 3 and November 5, 2026, Amazon's global MadPot decoy network intercepted requests sent to honeypots posing as large language model (LLM) inference APIs. Those requests inadvertently carried the agents' own operating memory and campaign logs.
How the self-funding key theft loop works
ARTEX pairs a Go backend and web console with standing instruction prompts that let autonomous agents manage reconnaissance, exploitation, and post-exploitation with minimal human oversight.
In the largest campaign observed by Amazon, operators instructed ARTEX agents to scan across a sequence of countries, prioritise enterprise production servers over personal projects, and grade every stolen credential by validity and quota depth. Keys exposed publicly for more than 30 days were flagged as low-value inventory.
The agents targeted self-hosted LLM gateways, proxies, and agent builders that store upstream provider keys centrally, including LiteLLM, Dify, One API, New API, Langflow, RAGFlow, Open WebUI, and Ollama. Once an agent extracted working API keys by exploiting uninitialised setups, default admin logins, or known flaws, it pooled those keys to pay for the LLM tokens needed to hunt for the next batch of victims.
KYC bypasses via deepfake injection
Amazon researchers also documented a separate ARTEX campaign aimed at cryptocurrency exchanges. In those operations, agents attempted to defeat know-your-customer (KYC) identity verification flows using real-time deepfake video injection.
Amazon notified potentially affected AWS customers referenced in the captured agent memory and published indicators of compromise so non-AWS operators can audit their logs.
Immediate steps for self-hosted AI teams
Teams running LiteLLM, Dify, One API, or local model gateways should remove admin consoles from public internet exposure immediately, place them behind a VPN or zero-trust proxy, rotate default master keys and provider API tokens, and apply current vendor patches.
Sources
- 1AWS Security Blog disclosureaws.amazon.com