November 16, 2026
Amazon Uncovers AI Agent Campaign Hunting Internet-Exposed LiteLLM, Dify and One API Deployments for Provider Keys
Amazon's threat intelligence team used MadPot decoy servers to capture records from attacker AI agents, likely running the open-source ARTEX pentest framework, that scanned internet-exposed AI apps and LLM gateways for stored provider keys. The playbook covered default admin credentials, unauthenticated management endpoints, an isCliRequest authentication bypass in One API family systems, SSRF to cloud metadata, and known Langflow and RAGFlow flaws. Stolen keys were pooled and reused to power further scanning.
LLM gateways and AI app builders are API-first systems concentrating many provider credentials behind admin endpoints. Discovering exposed AI APIs, enforcing authentication, and monitoring anomalous key usage is core API posture work.
Agentic/MCP
Read full story