Agent chaosVerified·18 Nov 2026
Hackers' AI agents stole AI keys from exposed gateways, then used them to pay for the hunt
Amazon's decoy servers caught attackers' AI agents raiding exposed LLM gateways and AI app builders for provider keys, then using the stolen keys to pay for the next round of hunting.
Amazon's threat intelligence team says attackers' AI agents have been raiding internet-exposed AI apps for the provider keys stored inside them. The best part: the agents told on themselves.
What happened
From 3 October to 5 November, Amazon ran decoy sensors, part of a network called MadPot, that pretended to be popular LLM inference APIs. Attackers' agents treated those decoys as the AI models powering their work and sent them prompts. Those prompts often carried the agents' own records of their campaigns, including operator instructions and what the agents said they had found.
Amazon assesses with moderate confidence that the agents ran on ARTEX, an open-source agentic penetration-testing framework, and links the main campaign, also with moderate confidence, to several separate China-nexus groups.
How the key hunt worked
The main campaign was built to pay for itself. Agents found exposed AI apps, pulled out the upstream provider keys, checked the keys worked and stored them in a shared pool. Those stolen keys then paid for the model calls that powered the search for more keys.
They went for apps that hold keys for many providers at once, because one weak admin page can expose many accounts. Targets included LLM gateways and proxies such as One API, New API, Veloera, Sub2API and LiteLLM, and AI app builders such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
The playbook was boringly effective: installs nobody had finished setting up, management endpoints with no login, default admin credentials such as LiteLLM's default master key, a header trick that bypasses authentication in One API family systems, known Langflow and RAGFlow flaws, and keys left in client-side JavaScript, source maps, debug routes and environment files. To find targets the agents also combed public code repositories, npm and PyPI packages, container images and paste sites.
Cloud credentials were treated as a door to more access, with agents going after managed AI model services and secrets stores. A second campaign used real-time deepfakes to try to beat identity checks at crypto exchanges.
The catch
Amazon did not watch these attacks happen and says it can't confirm what the agents claimed to achieve. What the records do show clearly is what the operators were after: keys that run up bills on someone else's AI accounts.
Our take
An LLM gateway is a keyring for every model account you have. If its admin page is on the internet with a default key, assume an agent has already found it. Take management interfaces off the public internet, change default master keys, finish or delete half-done installs, patch Langflow and RAGFlow, and rotate any provider key that ever sat somewhere public.
Treat an AI gateway like a password vault: never expose its admin side, never keep default keys, and rotate fast.

