Finding 1,000 exposed AI servers took researchers 10 minutes

Security researchers from Cisco's Talos team discovered over 1,100 exposed AI servers, specifically Ollama endpoints, using the Shodan scanning tool, with 20% actively hosting models vulnerable to unauthorized access. The rapid rise in large language model (LLM) adoption has often surpassed the establishment of robust security measures, leading to servers going online with default settings and inadequate protection. This scenario poses risks such as unauthorized model usage, excessive resource consumption, and potential exploits like model poisoning. Tailscale CEO Avery Pennarun highlighted the common oversight of placing private API servers on the public internet, advocating for better security practices and mentioning the Model Context Protocol's complexity as a frequently overlooked challenge. Despite this, solutions exist, such as using Tailscale for secure AI hosting, which emphasizes access control and network isolation without exposing systems to the broader internet.