CLOUDFLARE has announced new capabilities for its Model Context Protocol (MCP) to enhance security around AI agents. Traditional human-controlled resource permissions are at risk due to the rapid decision-making ability of AI agents, which can lead to numerous incorrect actions before human oversight. The new Cloudflare One features include traffic inspection for MCP, showing user actions and server interactions, and controlling direct connections on managed networks.
The MCP serves as a common access point for AI tools, exposing the potential for misuse if not properly monitored. The article elaborates on the anatomy of MCP tool calls, emphasizes the importance of security check sites within the network, and categorizes the three control layers available for MCP requests: client-side, network boundary, and server-side.
It highlights the introduction of a dashboard for visibility into MCP traffic, improved detection heuristics for classifying MCP requests in Gateway, and the option to enforce strict access controls via cloud gateways.