Home · Free tools · MCP Security Baseline

Free tool · built by Korelex

Consumer MCP Security Baseline Public draft v1.0

Before you let ChatGPT into your customers' accounts.

Security requirements for a remote MCP server that customers connect to their own accounts through an AI client such as ChatGPT, Claude or Gemini — with a self-assessment that tells you whether you are ready to launch.

What it covers

Six domains, one outcome each.

  • AData & PrivacyCustomer data stays with the customer it belongs to. 4 controls.
  • BSecurityOnly the real user, through an approved client, gets in. 8 controls.
  • CSafetyHigh-impact actions need a human, outside the chat. 2 controls.
  • DReliabilityTool calls do exactly what the user asked, once. 2 controls.
  • EAccountabilitySomeone owns it, sees problems and can act fast. 6 controls.
  • FSocietyYour platform can't be turned into a tool for abuse. 2 controls.

What you get

A launch verdict, and what to fix first.

Answer each control and the assessment returns one of four verdicts. The rules run in order and the first that applies decides: any launch blocker missing or unknown means Not ready.

  • Not ready
  • Partially ready
  • Ready, pending verification
  • Ready
  • What to fix, in order. Launch blockers to close before real customers connect, then the controls to close within 30 days of launch.
  • Attacks still open. The attack paths your gaps leave open.
  • Reporting obligations to plan for. What you would have to report, and to whom, if something goes wrong.
  • A report you can keep. Save it as a PDF for your team.

How it is built

Every control can be tested.

Each control has one requirement statement, an objective, where it applies, the evidence required and a test procedure: what to do, and what you should see. Mappings to other frameworks are for alignment only.

  • OWASP Top 10 for LLM Applications 2025
  • OWASP API Security Top 10 2023
  • ISO/IEC 27002
  • MCP Security Best Practices
  • AIUC-1 domains

KX-MCP-A1 · launch blocker

Ownership checked on every call

Requirement
The MCP server SHALL authorize every tool call against the user identified by the access token, for every record the call reads or changes.
Test
With user A's token, call every tool using user B's account, order and record IDs. Every call is refused with 403 or not found.
Maps to
OWASP API1:2023 Broken Object Level Authorization; ISO/IEC 27002 8.3 Information access restriction.

Scope

For servers your customers connect to.

In scope

  • The MCP server, its tools and the responses they return
  • The authorization server and the flow that connects an AI client to a customer account
  • Your processes around the server: ownership, monitoring, incident response and change control

Out of scope

  • The AI client and the model behind it
  • MCP servers that run locally on a user's device
  • MCP servers used only by your own staff
  • Your wider platform, which ISO/IEC 27001 and SOC 2 already address

This is a public draft for comment: requirements, IDs and mappings may change before version 1.0 is final. It is not an AIUC-1 audit or certification, and Korelex is not affiliated with AIUC.

When belief is not enough

Self-assessment tells you what you believe. Testing tells you what holds.

An AI Red Team attacks the running server — the account-linking flow, cross-customer isolation, prompt injection through tool results — and reports only what we could make happen. See how we test MCP servers.

Pricing shared on request.