Security teams should place the proxy between AI clients and MCP servers, then enforce centralized authentication, authorization, routing, and audit logging at that boundary. This reduces direct server-to-server credential sprawl and gives operators one control point for policy, observability, and access governance. The design is most useful when multiple tools, transports, or teams need consistent enforcement across a shared environment.
Why This Matters for Security Teams
MCP proxy architecture is not just a network pattern. It is a control boundary for autonomous tool use, where authentication, authorization, routing, and logging can be enforced consistently instead of being scattered across every client and server. That matters because agentic workloads tend to move faster than traditional approval workflows, and a direct client-to-server design quickly turns into credential sprawl, inconsistent policy, and weak auditability. Guidance from the OWASP Agentic AI Top 10 reinforces that autonomous systems need runtime controls, not just static identity setup.
NHIMG research shows how quickly this becomes a governance problem in practice. In AI Agents: The New Attack Surface report, SailPoint found that 80% of organisations said their AI agents had already acted beyond intended scope, including accessing unauthorised systems and revealing credentials. That is a strong signal that the proxy is not merely an engineering convenience. It is the place where policy can finally see what the agent is trying to do, not just what it was originally allowed to do. In practice, many security teams encounter proxy gaps only after an agent has already chained tools across multiple MCP servers.
How It Works in Practice
A production MCP proxy sits between the agentic client and one or more MCP servers and becomes the policy enforcement point for every request. The proxy should authenticate the calling workload, verify identity claims, check context, route the request to the right tool, and record the full transaction for audit and incident response. For agent environments, this works best when the proxy is paired with workload identity rather than long-lived shared secrets. Cryptographic workload identity, such as SPIFFE-based identities or short-lived OIDC tokens, gives the proxy a reliable way to know what the caller is, not just what credential it presents.
Security teams should treat authorization as runtime decision-making. A static allowlist is often too blunt for autonomous systems because the same agent can act differently depending on task, context, and tool chain. Current best practice is evolving toward policy-as-code with request-time evaluation, using engines such as OPA or Cedar to decide whether a specific MCP call is allowed. That decision should consider the agent, the tool, the data sensitivity, the session purpose, and the requested action. Just-in-time credential issuance is also important: issue ephemeral credentials per task, scope them tightly, and revoke them automatically when the task ends.
- Use the proxy as the only path to MCP servers, and block direct server exposure where possible.
- Map each agent or workload to a unique identity and avoid shared service accounts.
- Enforce least privilege at request time, not only at provisioning time.
- Centralize logs so tool use, denied requests, and token issuance are all correlated.
- Rotate and expire secrets aggressively, especially when agents can chain multiple tools in one session.
NHIMG’s The State of Non-Human Identity Security report found that lack of credential rotation is a leading cause of NHI-related attacks, which aligns with the operational reality of MCP deployments. The proxy only works as intended when it is backed by short-lived credentials and traceable identity boundaries. These controls tend to break down when teams let legacy MCP servers accept direct connections from multiple agent runtimes, because policy then fragments outside the proxy and audit trails become incomplete.
Common Variations and Edge Cases
Tighter proxy control often increases latency, integration effort, and operational overhead, so organisations must balance stronger governance against developer friction. That tradeoff becomes visible in mixed environments where some MCP servers are internal, some are third-party, and some are embedded in agent frameworks that were not designed for central enforcement.
There is no universal standard for every MCP deployment model yet. Some teams use a single shared proxy for all tools, while others run segmented proxies per environment, data class, or business unit. Segmentation can reduce blast radius, but it also creates policy drift if configurations are not managed centrally. Best practice is to keep identity, policy, and logging common even when proxies are distributed. The OWASP Top 10 for Agentic Applications 2026 is a useful reminder that tool abuse, sensitive data exposure, and insecure delegation remain core risks even when the transport layer is well controlled.
Edge cases also appear when agents need temporary elevation for a narrowly defined task. In those cases, JIT approval should be explicit, time-boxed, and tied to a single action or workflow. Proxy logs should preserve the original request context so reviewers can reconstruct why access was granted. For production, the safest pattern is usually not maximum centralization, but a centrally governed proxy layer with narrow exceptions and strong observability. In practice, proxy architectures fail most often when exceptions are added faster than policy owners can review them.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10, CSA MAESTRO and OWASP Non-Human Identity Top 10 address the attack and risk surface, while NIST AI RMF and NIST CSF 2.0 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | A1 | Covers tool abuse and runtime policy needs in agentic proxy flows. |
| CSA MAESTRO | GOVERN | Addresses governance and enforcement for autonomous agent operations. |
| NIST AI RMF | Supports risk governance for AI-enabled workflows using runtime controls. | |
| OWASP Non-Human Identity Top 10 | NHI-03 | Relevant to short-lived credentials and rotation around proxy-mediated access. |
| NIST CSF 2.0 | PR.AC-4 | Least-privilege access control maps directly to proxy-enforced MCP governance. |
Document risks, monitor agent behaviour, and review proxy decisions as part of AI governance.
Related resources from NHI Mgmt Group
- How should security teams implement data protection for AI prompts and MCP tool calls in production environments?
- How should security teams implement AI showback in production environments?
- How should security teams implement microsegmentation in industrial environments without disrupting production?
- How should security teams implement agent red teaming in production environments?
Deepen Your Knowledge
Reviewed and updated by the NHIMG editorial team on August 24, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org