AI Servers Silently Hijacked Through Weak Configurations

An investigation revealed that thousands of servers running AI inference platforms — such as Ollama, vLLM, and local model servers — were exposed to the internet without authentication, silently turned into cryptocurrency miners and entry points into corporate networks. ## The problem: open ports without protection Tools like Ollama, by default, listen on port 11434 on all network interfaces. When deployed on cloud servers without proper firewall configuration, they become publicly accessible. Attackers scanning the internet with tools like Shodan or Censys find these instances within minutes. ## What attackers do Once with access to the AI server, attackers can install crypto miners leveraging high-performance GPUs, use the server as a proxy for attacks, exfiltrate proprietary models, access data sent for inference by other users, and pivot into the company's internal network. ## How to protect your AI servers 1. Never expose inference APIs directly to the internet 2. Use a reverse proxy (Nginx, Caddy) with authentication 3. Configure firewalls to allow access only from trusted IPs 4. Implement GPU usage monitoring to detect mining 5. Enable logs and unusual access alerts 6. Use VPN or private networks for remote access