⚠ Enterprise Security

Securing API, Streaming, MCP & LLM Traffic
Inside Your Enterprise

AI agents talking to MCP servers — and to LLM providers like Anthropic and self-hosted Ollama — create distinct threat vectors most enterprises haven't planned for. Here's how to think about them — and how to close the gaps.

Agent Galaxy — principal access graph showing entitlements versus observed usage across users, agents, and MCP tools
Agent Galaxy — see every principal's entitled-vs-actual access to MCP tools and LLM models, live.
01
The Rogue Server Problem High risk
What enterprises don't know is already running
⌄

Right now, developers inside your enterprise are spinning up MCP servers in places nobody is tracking — on laptops, in personal cloud accounts, as VS Code extensions, in random Docker containers. These servers expose tools: file system access, database queries, API calls, internal service integrations.

An AI agent pointed at one of these servers can execute actions — not just read data. This is fundamentally different from traditional shadow IT.

⚠

The blast radius problem: A rogue SaaS app accesses data. A rogue MCP server lets an AI agent execute actions at scale — file writes, API calls, database mutations — with no audit trail and no governance.

The platform doesn't automatically discover rogue servers outside it. What it does is give you a governed chokepoint — a single gateway that all approved MCP and LLM provider traffic flows through, combined with network controls to enforce that nothing bypasses it.

02
North-South Traffic External agents
AI agents outside your perimeter reaching internal MCP servers
⌄

North-south traffic is what most people think about first — an external AI agent trying to reach an MCP server inside your perimeter, or an internal MCP server or agent calling out to unauthorized external services, including LLM provider APIs.

Uncontrolled
External AI Agent → Firewall → Any MCP Server (direct)
Internal MCP Server → Firewall → External APIs (unrestricted)
With platform
External AI Agent → Firewall → Envoy Gateway → Approved MCP Servers only
Direct MCP access → Blocked at firewall

How to enforce it

Your network team configures the perimeter firewall to allow MCP traffic only to and from the platform's Envoy gateway IP. All other MCP traffic is dropped. Same pattern enterprises use for SaaS allowlists.

  • All external AI agents point at your gateway hostname (e.g. mcp.company.com)
  • Firewall blocks all other inbound MCP traffic
  • Egress filtering prevents MCP servers and agents calling unauthorized external APIs — including LLM providers outside your approved list
  • Every request is logged, governed, and auditable
03
East-West Traffic Harder problem
Internal agents bypassing your perimeter entirely
⌄

East-west traffic never crosses your perimeter. A developer's AI agent on their laptop talks directly to an internal MCP server — bypassing the firewall, bypassing the gateway, bypassing every control you have.

⚠

Perimeter controls don't help here. The traffic is entirely internal — it looks like normal developer activity from a network perspective.

East-West threat
Developer's AI Agent (laptop) → Internal network → Rogue MCP Server (10.0.1.50:3000)
Never touches the gateway. Never logged. Never governed.
East-West controlled
Internal AI Agent → Network policy routes through gateway → Approved MCP Servers only
Direct pod-to-pod MCP access → Blocked by Kubernetes NetworkPolicy

Why it's the harder problem

Internal MCP servers and LLM gateways often have access to the most sensitive resources — databases, private APIs, file systems, messaging platforms, and the prompts and completions flowing to model providers. An AI agent with access to an ungoverned internal server can exfiltrate data and mutate state entirely within your network perimeter.

How to enforce it

  • Internal network proxy routes all MCP traffic through the gateway
  • AI agent configurations managed centrally — all agents point to the gateway
  • Kubernetes NetworkPolicy blocks direct pod-to-pod MCP connections
  • Port-level controls block direct access to MCP server ports from non-gateway sources
ℹ

Layer 3 is automatic. Kubernetes NetworkPolicy ships with every installation — MCP pods reject connections from any source except the Envoy gateway. Layers 1 and 2 require your network team, and we provide the reference architecture.

Enforcement Model

Three layers of control

Complete MCP and LLM traffic governance requires enforcement at three distinct layers. The platform handles Layer 3 automatically — Layers 1 and 2 are implemented by your network team using standard enterprise tooling.

North-South — block everything except the gateway

Allow inbound MCP traffic only to the platform's Envoy gateway IP. Block all other MCP traffic at the perimeter. Configure egress filtering to prevent MCP servers and agents from calling unauthorized external services, including LLM provider endpoints outside your approved list.

  • Add allowlist rule for gateway IP on HTTPS/443
  • Drop all other inbound MCP traffic at the firewall
  • Configure egress filtering for MCP server pods
  • Same pattern enterprises use for SaaS allowlisting
Authorization Model

Both principals, or neither

Most gateways authorize the agent. It presents a service-account token, the token carries scopes, the call goes through. That model holds until one agent starts serving many people — at which point the service account becomes every user’s ceiling.

Magertron evaluates both principals on every call. The agent must hold the permission, and so must the human it is acting for. The effective grant is the intersection of the two, never the union.

Agent grants service account Denied Human grants invoking user Denied Allowed ∩ Union here would be the vulnerability.

Close these gaps in your own Kubernetes cluster.

Magertron runs in your cluster, not ours. No data leaves your perimeter. Free up to 20 servers — no signup, no credit card. Try it in 5 minutes.

Try it now → View on GitHub →