Skip to content

Security Review

Security

Security review for MCP Gateway is a control-by-control check of the governed MCP path. The reviewer should be able to prove identity, default deny, discovery filtering, selected-operation boundaries, credential secrecy, connector routing, session revocation, metadata-only audit, customer-controlled telemetry, and release evidence without trusting marketing claims.

  • Security reviewers approving a pilot or release candidate.
  • Platform owners preparing hardening evidence.
  • Docs authors checking whether a public security claim is sourced.
  • Operators responding to incident-control findings.

This page is the security-review checklist for public docs. It is deeper than the Security Model summary: each control maps to a source-backed verification path and a failure action. If a behavior is not sourced, mark it needs-source instead of smoothing it into public prose.

Use it before publishing security docs, adding generated references, changing release evidence, or presenting deployment hardening. Use it again when a reviewer asks for a deferred capability such as full DLP, SCIM/device posture, tamper-evident audit hash chains, local STDIO governance beyond approved adapter profiles, or managed MCP hosting.

  1. Confirm production auth mode and actor normalization.
  2. Prove Cedar default deny and explicit deny overriding allow.
  3. Prove unauthorized tools are hidden during discovery, not only denied at execution.
  4. Prove API-backed tools come only from selected and approved operations with host, schema, method/path, credential, timeout, and size checks before upstream.
  5. Prove credential broker output is opaque handles or safe refs, never secret material.
  6. Prove connector disable or revoke blocks routes before upstream traffic.
  7. Prove session revocation blocks new calls and affects active sessions according to the contract.
  8. Prove audit/export, telemetry, release bundles, screenshots, docs, fixtures, and PR text are metadata-only.
ControlWhat the reviewer checksFailure action
IdentityOIDC/JWT, trusted proxy, SAML-through-proxy, workload identity, or local identity produces normalized actor context.Deny discovery and calls until identity refs validate.
PolicyCedar simulation shows default deny and forbid precedence.Preserve deny and version the policy fix.
DiscoveryUnauthorized tools are hidden from catalog/discovery.Treat visible unauthorized tools as a release blocker.
CredentialsBroker resolves allowed modes without returning material.Disable or revoke the binding and remove unsafe output.
ConnectorsRegistered private routes only; disabled or revoked routes fail closed.Block new routes and preserve safe audit.
SessionsDrain, terminate, revoke, expiry, and audit follow explicit session states.Reject new stateful sessions or revoke affected active sessions.
Audit/exportEvents carry policy version, credential mode, request/session IDs, reason code, redaction status.Remove payload-bearing evidence and rerun no-secret gate.
Release evidenceCosign, SBOM, provenance, backup/restore, upgrade/drain, rollback, checksums, no-secret scan.Do not publish the release candidate until evidence validates.
  • A public page claims ML prompt-injection detection, full DLP, SCIM/device posture, endpoint command sandboxing, tamper-evident audit hash chains, local STDIO governance outside approved adapter profiles, or managed MCP hosting as current V1 behavior. Those need source or must stay deferred/outside V1.
  • A security finding is closed using a screenshot or workpad instead of source contracts, runbooks, schemas, or harness checks.
  • A diagnostic includes backend details, payloads, prompts, responses, tokens, credential material, or customer data.
  • A connector, credential, policy, or session revocation does not fail closed before upstream execution.
  • Release artifacts include private material or generated local outputs that were never curated for public release.

Type set in Geist, Source Serif 4, and Departure Mono.