Exposed AI Supply Chain Endpoints Present Security Risks

• •36,769 AI endpoints exposed; only 2.02% have authentication.
• •Ollama servers can be accessed anonymously, risking resource abuse.
• •Lack of network perimeter security is a widespread issue.

A recent study revealed 36,769 self-hosted AI endpoints are exposed on the internet, with only 2.02% protected by HTTP authentication. The research conducted by Mysterium VPN indicates that running AI locally does not guarantee security, as many infrastructures are publicly accessible. The most concerning findings include 18,529 Open WebUI instances, with only one behind an authentication gate, and 6,935 Ollama servers that respond to anonymous requests. These exposed endpoints can lead to unauthorized access and resource abuse, termed 'LLMjacking.' The report highlights a significant lack of network perimeter security across various AI applications, raising alarms about potential exploitation. The findings suggest a broader issue in the AI supply chain security landscape, with implications for organizations relying on self-hosted AI solutions.