MCP & tool security

A gateway in your cloudfor every tool call.

A gateway in your VPC sits in front of every MCP server, model and API your agents call. Payloads never leave your environment.

The problem

Every tool is
a way in.

MCP made it easy to give agents tools. It also gave attackers a place to hide instructions, widen access or pull data out.

  • Poisoned toolsA tool description or result carries instructions the agent follows.
  • Over-scoped keysOne MCP server key reaches far more than any task needs.
  • Data leaving your cloudPayloads sent to tools and models outside your environment.
A real pathA tool update tries to move money.
  1. 1member-chat-assistThe agent calls kb.search through the gateway in your VPC.
  2. 2mcp kb-syncA new version of the server adds an instruction to call accounts.transfer.
  3. 3gatewayThe call is checked against the agent’s grant, inside your cloud. Transfers are not in it.
  4. BlockTool not in grant. The call never reaches core banking.mcp core-banking · accounts.transfer
Industry references
  • ASI04Agentic Supply Chain VulnerabilitiesOWASP Top 10 for Agentic Apps
  • ASI02Tool Misuse and ExploitationOWASP Top 10 for Agentic Apps
  • AML.T0110AI Agent Tool PoisoningMITRE ATLAS
How Zenstra handles it

Inspect, scope and stop
inside your own cloud.

The gateway is the decision point for every call your agents make to tools, models and APIs.

Inspect

Every call, inside your VPC

The gateway sees each MCP, model and API call before it leaves the agent, without sending prompts or payloads anywhere else.

  • Deployed with Helm or Terraform
  • Runs in AWS, Azure or Google Cloud
  • Only decisions and redacted evidence reach the control plane
Scope

Each tool becomes a grant

A tool call is checked against the agent’s grant: which tool, which scope, for how long.

  • Per-agent identity for every call
  • Short-lived grants for model providers
  • Unapproved MCP servers flagged
Stop

Blocked before it lands

Calls outside the grant, poisoned tool content and data leaving for unapproved destinations are stopped inline.

  • Decided inline, before the call lands
  • Fails closed if unreachable
  • Every block on the decision record
Policy as code

The policy
behind it.

Every MCP server, model and API sits behind one gateway. Each tool is a grant, and instructions hidden in tool content are quarantined.

  • Decided inline
  • Fails closed
  • Versioned with every decision
denymcp.callwhen tool not in grantdenytool.resultwhen it carries instructions · quarantineholdmcp.serverwhen new or changedallowkb.searchfor member-faq
What it stops

The attacks that
come through tools.

  • Tool poisoning

    Instructions hidden in tool descriptions and results.

  • Over-scoped calls

    A transfer when the task only allowed a read.

  • Rogue MCP servers

    Tool servers nobody approved, found and blocked.

  • Data exfiltration

    Payloads headed for destinations outside policy.

  • Credential misuse

    Standing keys replaced by per-agent, expiring grants.

  • Model sprawl

    Calls to unapproved models and endpoints.

FAQ

Common questions.

Where does the gateway run?

In your environment: your VPC or Kubernetes cluster on AWS, Azure or Google Cloud, deployed with Helm or Terraform.

Does any data leave our cloud?

Prompts and payloads are evaluated inside your environment. Only decisions and redacted evidence reach the Zenstra control plane.

Does it work with MCP servers we did not build?

Yes. The gateway sits in front of the servers your agents call, whoever built them.