New updates have been reported about OX Security.
OX Security has published a large-scale analysis of 15,465 Model Context Protocol (MCP) servers, warning that AI agents are increasingly connecting to infrastructure that sits outside traditional cloud security and governance controls. Based on 5,095 unique hostnames, the company found that 15.6% resolved to infrastructure outside the U.S., including assets in China and Russia, as well as home networks, locally hosted machines and even abandoned domains that can be re-registered cheaply.
The company’s researchers concluded that MCP is effectively creating a new, weakly governed layer of AI infrastructure, where organizations often lack visibility into who operates servers, what data they receive or whether the running code matches published source. OX Security’s testing also showed how a malicious MCP server can escalate a single “always allow” approval into unauthorized access to sensitive files via prompt injection, underscoring that AI agents and their reachable infrastructure are now central to enterprise security posture and must be formally governed alongside existing cloud environments.