MCP Security

MCP security is about controlling what AI tools can read, write, and execute.

Simple idea: AI tool access must be permissioned, logged, and limited.

Why It Matters

An AI assistant connected to tools can be powerful, but also risky. If permissions are too broad, it may:

  • Read private files.
  • Leak sensitive data.
  • Modify the wrong document.
  • Call the wrong API.
  • Take actions without approval.
  • Expose customer or company data.

Security Principles

PrincipleMeaning
Least privilegeGive only the access needed
Human approvalRequire approval for risky writes or external actions
Audit logsRecord every tool call
Data boundariesSeparate public, internal, confidential, and customer data
Tool allowlistOnly expose trusted tools
Input validationDo not trust model-generated arguments blindly
Output filteringAvoid returning secrets or private data unnecessarily

Risky Tool Types

  • File deletion.
  • Database writes.
  • Email sending.
  • Payment or billing actions.
  • Customer-data exports.
  • Production deployment actions.
  • Admin-permission changes.

Safe Defaults

  • Read-only tools first.
  • Approval for writes.
  • No destructive actions by default.
  • Clear user-visible logs.
  • Per-vault or per-folder access.
  • Per-environment separation between dev and production.

CookieYes Example

An AI assistant that reviews consent setup should be able to read configuration and scanner output, but should require approval before changing live consent rules.

Mozilor Example

An e-commerce agent can draft a refund or inventory action, but should require approval before issuing refunds or changing prices.