In little more than a year, the Model Context Protocol (MCP) went from proposal to the standard way of connecting AI agents to tools, data and systems. Nearly every AI platform supports it, and more and more software vendors ship MCP servers for their products.
That is good news: a common standard keeps every integration from being a custom build. But the speed of adoption came at a cost, and this year it became visible.
The signals
The signals come from several directions. Security researchers found thousands of MCP servers exposed on the internet, many with no authentication at all. Recorded vulnerabilities mentioning MCP number in the hundreds, the vast majority published in 2026. In May, the NSA published secure design guidance for MCP whose core finding is plain: adoption has outpaced the security model.
On top of that comes a deeper change. The July revision of the specification made the protocol stateless. That helps it scale but, as Akamai analysts warned, it shifts security decisions the protocol used to handle onto each implementation.
And there is a major gap: MCP does not define a role-based access control model. Mapping each request to a verifiable identity and specific permissions is left to each company.
My view
MCP is not insecure in itself. The problem is how it is being adopted: as if connecting an agent to a system were just another integration, when in fact it means giving autonomous software the ability to act on that system.
In banking, insurance or any critical environment, that requires the same rigor as any other privileged access. Today, in many organizations, it does not get it.
Connecting agents without opening a back door
These are the controls I consider the minimum before connecting an agent to a system that matters:
1. No MCP server exposed without authentication. It sounds obvious, but the data shows it is not. Use OAuth with short-lived, narrowly scoped tokens, never static credentials embedded in configuration.
2. Separate agent identity from user identity. The system must know on whose behalf the agent acts and also which agent it is. Effective permissions should be the intersection of both, never the sum.
3. A gateway as the control point. Instead of every agent connecting directly to every server, a gateway centralizes authentication, authorization, logging and rate limits. It is the same pattern used for years with APIs, applied to agents.
4. Least-privilege tools. An MCP server that exposes “run any query” is very different from one that exposes “check this customer’s balance.” Designing narrow tools is the most effective defense against misuse, including misuse induced by malicious instructions hidden in data.
5. Inventory and review third-party servers. Every third-party MCP server is code running with access to your systems. It deserves the same scrutiny as any other critical dependency.
6. Full logs of what each agent did. Which tool it called, with which parameters, on whose behalf and with what result. Without that, there is no audit and no incident response.
Conclusion
MCP will remain the foundation of how agents integrate with enterprises, and rightly so. The question is not whether to adopt it but with which security architecture. Organizations that define it now will be able to connect agents to critical systems with confidence. The rest will discover their weak points once they are already in production.
I help organizations design the identity, security and integration architecture to connect AI agents to their systems. See services.
Sources: SecurityWeek on the new MCP specification; Cloud Security Alliance MCP security best practices.
