The AI Supply Chain Has a Security Problem, and Much of It Is Sitting on the Open Internet

A Mysterium VPN study found at least 36,769 self‑hosted AI endpoints reachable from the public internet—model servers, agent builders and vector stores—of which only about 2% presented an HTTP authentication challenge. The report highlights large populations of exposed Open WebUI and Ollama instances, numerous agent/workflow platforms (Flowise, n8n, Langflow, etc.) that may hold API keys and credentials, and vector stores that could contain sensitive data; researchers intentionally measured exposure (not compromise) and recommend treating self‑hosted AI infrastructure like any other internet‑facing service with proper network controls and authentication.