This is because traditional permission models were designed for humans. A senior engineer can deploy code and query databases, but the risk has limits—humans exercise judgment, and humans have speed limits. AI Agents change the rules: their decisions can be uncertain, and they can repeat the same action without bound.
We've spotted a more practical problem: employees can, with a single configuration line, connect Claude Code, Cursor, or VS Code to a company's MCP server without IT approval. Cloudflare calls these approval-bypassing connections "Shadow MCP," and this week's new release targets identifying and intercepting that traffic at the network layer.
What This Is
MCP (Model Context Protocol) is an open standard launched by Anthropic last year that gives AI Agents a common language for connecting to third-party tools, databases, and internal company systems. It's like giving AI a pair of "hands" to do work directly.Cloudflare's solution operates on three layers: the client, the network layer, and the server itself. Its Gateway product can identify MCP traffic signals at the network layer and force all connections through approved paths.
Industry View
Supporters call this "security infrastructure for the AI era." Anthropic, OpenAI, and Microsoft have all integrated MCP into mainstream development tools. The enterprise data exposure surface is now orders of magnitude larger than before—someone has to catch what falls through.But there are dissenting voices. Security researchers note that the MCP protocol itself doesn't specify hostname or path signatures—any HTTPS call could be MCP traffic, making pure network-layer identification imprecise. Consolidating traffic through a single gateway also creates a new single point of failure and attack surface. A European bank security chief told us privately: "What worries us more is Agents learning to bypass it."
Impact on Regular People
For enterprise IT: the traditional "perimeter defense" mindset is broken. Agents have no fixed IP, no fixed schedule—IT teams can no longer troubleshoot by asking "who connected to what, and when."For white-collar workers: AI tool usage will face increasing restrictions. Compliance departments will require approved tools; unapproved Cursor or Claude Code installations may be blocked by employers.
For the consumer market: the cybersecurity industry landscape may reshuffle. Established players like Cloudflare, Zscaler, and Palo Alto Networks are all scrambling for the new prize of "AI Agent traffic."