MCP Security Checklist

Who the agent is, and what it is allowed to do on whose behalf.

• Each MCP tool runs with least-privilege credentials, not a shared admin token.
• The agent's identity is distinct from the end user's — no ambient or inherited authority. Prevents confused-deputy situations where the agent acts with more privilege than the requesting user holds.
• Access scopes are defined and enforced per tool, not granted broadly across the server.
• Credentials and tokens are short-lived and rotated; long-lived static secrets are avoided.

What each tool can be asked to do, and with what inputs.

• Tool parameters are validated against a strict schema before execution.
• Tools cannot be invoked with arbitrary paths, queries, or commands supplied by the model.
• The tool catalog exposed to the model is the minimum needed for the task. A smaller surface means fewer ways for a manipulated model to take an unintended action.
• High-impact tools (delete, pay, send, deploy) are clearly separated and gated.

Where a person stays in the loop — especially for irreversible actions.

• Consequential actions (writes, payments, deletions, external sends) require explicit human approval.
• Approval prompts show the actual action and parameters — not a model-written summary. A summary can be manipulated; the real call cannot hide what it will do.
• Users can deny or modify a proposed action before it executes.

Treating retrieved data and tool output as untrusted by default.

• Retrieved content and tool output are treated as untrusted input (indirect prompt injection).
• System instructions are isolated from user- and tool-supplied content.
• Untrusted content cannot silently change which tools are called or with what arguments.

Being able to see, after the fact, exactly what happened.

• Every tool invocation is logged with inputs, outputs, and the acting identity.
• Logs are tamper-resistant and retained according to policy.
• Anomalous tool-use patterns are monitored and can raise alerts.

Where the MCP servers come from and what they can reach.

• MCP servers come from trusted sources and are pinned to known, reviewed versions.
• Network egress from tools is restricted to the destinations they actually need.
• Secrets are never passed through the model context or prompt.