Amazon Threat Intelligence says AI agents ran a large-scale campaign to steal API keys and other credentials for AI models from internet-exposed applications. The operation targeted LLM gateways and proxies including LiteLLM, One API, New API, Veloera and Sub2API, as well as AI application builders such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
Amazon's MadPot decoy network captured the agents' requests and operational records between October 3 and November 5, 2026. The sensors posed as popular LLM inference APIs. Amazon assesses with moderate confidence that the activity used ARTEX, an open-source agentic penetration-testing framework.
According to the captured records, the agents located exposed AI applications, extracted upstream provider keys, tested whether they worked and placed them in a shared pool. Stolen keys were then fed into the framework's own model gateway, paying for the inference used to hunt for more credentials. The records describe a playbook covering default administrator credentials, unfinished installations, unauthenticated management endpoints, known software vulnerabilities and secrets exposed in code or deployment artefacts.
Amazon cautions that its sensors did not directly observe the activity against other organisations and that it cannot verify all outcomes reported by the agents. The company says it notified potentially affected AWS customers where appropriate.
For organisations running LLM infrastructure, Amazon recommends treating AI applications, model gateways and inference endpoints as production systems. Its advice includes replacing default credentials, completing installations, patching vulnerable software, removing and rotating exposed secrets, and monitoring unexpected model usage or spending. Amazon has also published indicators of compromise to help defenders check their environments.