MCP Implementation Security Review

MCP Implementation Security Review

Process

Step 1 — Classify the target

• Check MCP protocol version 2025-03-26 or later (current: 2025-11-25). Flag older versions as a finding but continue the review.
• Determine whether the target is a server or client.
• Classify transport as network-exposed or local-only using the transport reference below.
• Record transport, protocol version, and whether sessions exist.

Completion criterion: Target type, protocol status, and transport are identified.

Step 2 — Filter false positives

• Apply the False Positive Filters before opening findings.
• Keep docs only when they describe the repo's own server behavior, deployment, transport, or auth posture.
• For framework/SDK repositories, scope findings to the default configuration and public API surface.

Completion criterion: Remaining evidence is in-scope code, repo-owned docs, or public API behavior.

Step 3 — Check baseline controls

• For network-exposed servers, check MCP-01 through MCP-05.
• For local/STDIO servers, do not mark baseline controls PASS/FAIL; give best-practice notes and continue to RCE review.
• For clients, only review token/session handling explicitly visible in client code; do not apply the server baseline unless the user asks for client-side risk review.

Completion criterion: Each applicable control has a supported status.

Step 4 — Check RCE vectors

• Review all 7 RCE vectors.
• Mark each vector SAFE, AT RISK, or N/A.
• Prefer direct evidence over inference; the RCE Vectors table below enumerates the patterns to look for.

Completion criterion: Every relevant tool has an RCE result or explicit N/A.

Step 5 — Check OWASP MCP Top 10

• Evaluate all 10 OWASP risks below.
• If a control from Step 3 already fully covers an OWASP risk, reference that result rather than re-checking.
• For local/STDIO servers, mark network-dependent OWASP risks (MCP07, MCP09) as N/A.
• Mark each risk PASS, FAIL, or NEEDS INVESTIGATION.

Completion criterion: All 10 OWASP risks have outcomes supported by observable evidence or referenced from Step 3.

Step 6 — Report

• Use the Compliance Output Format below.
• Include file/line references in every justification.
• Separate code findings from manual follow-ups.
• If evidence is incomplete, use NEEDS INVESTIGATION and name the missing artifact.

Completion criterion: The report includes controls, RCE, optional OWASP, and actions.

Reference

Decision rules

• Network-exposed server: Apply all 5 controls, then run RCE and requested OWASP checks.
• Local/STDIO server: Give best-practice guidance only for the 5 controls; still run RCE because tool input can execute locally.
• Client: Review received-token handling and refusal to trust server-provided session IDs; do not force server controls unless asked.
• Reverse proxy or container exposure: If traffic can reach the server over a network, treat it as network-exposed even if inner binding is localhost.
• Unclear evidence: Do not guess. Mark NEEDS INVESTIGATION and say what must be verified manually.
• Ambiguous auth coverage: Auth middleware exists but it is unclear whether it covers MCP endpoints → mark NEEDS INVESTIGATION.
• Undeterminable transport: If transport cannot be established from code, flag for manual review and do not assume STDIO — defaulting to STDIO would wrongly skip the server controls.

Transport classification

Network-exposed (enforce all controls):

Local-only (best practices only):

Host binding gotchas:

False Positive Filters

Docs describing the repo's own server behavior, transport, auth posture, or deployment are not false positives.

Controls Reference

MCP-01 — Identity isolation

Scope: Remote MCP servers

Condition

• Authenticate every inbound request with a trusted identity provider and enforce authorization at the server boundary; do not infer auth from session IDs, prior requests, or network location.
• Use a unique server-specific application identity and audience/resource identifier; outbound calls use independently scoped service credentials or on-behalf-of flow where required, never the inbound token.
• Unauthenticated discovery endpoints are allowed only for metadata-only OAuth/MCP bootstrapping: /.well-known/oauth-protected-resource, /.well-known/oauth-authorization-server, /.well-known/openid-configuration.

What to check

• Token validation and authorization middleware run on every MCP route; authorization distinguishes tool invoke, read-only, and admin operations if present.
• Identity config shows a dedicated application/client/resource ID and audience; outbound clients acquire their own tokens and never copy inbound Authorization.
• Discovery endpoints return metadata only and cannot execute tools or expose protected data.

Key pitfall: Shared application identities or forwarded caller tokens break identity isolation and create confused-deputy paths.

MCP-02 — Sessions

Scope: Remote MCP servers that support sessions

Applicability

• No session identifiers issued or used anywhere → mark N/A (per-request auth is still required; see MCP-01).
• Sessions managed by the transport/SDK (e.g., Streamable HTTP Mcp-Session-Id) but generation/binding not visible in source → mark NEEDS INVESTIGATION, not FAIL.
• Session identifiers present in code → score PASS/FAIL against the conditions below.

Condition

• Authenticate and authorize every request; session state never substitutes for token validation.
• Session IDs are opaque correlation/continuity tokens only; they do not grant privileges, encode authorization, or bypass auth.
• Session IDs are CSPRNG-generated, unpredictable, bound to an authenticated context, and never embedded in URLs.

What to check

• Middleware validates tokens per request, not only when a session starts.
• Authorization logic never trusts a session ID alone; loss or reuse of a session ID must not grant access.
• Session creation uses random IDs (GUID v4/CSPRNG acceptable; sequential or time-based IDs are not).

Key pitfall: Treating a session ID as a bearer credential turns a correlation token into authentication.

MCP-03 — Rate limits

Scope: MCP servers and tools

Condition

• Enforce rate limits and abuse protection on tool discovery and tool invocation.
• Enforce limits at the MCP server runtime, not only at a gateway; partition by authenticated identity and by session where sessions exist.
• Apply stricter limits to mutation-capable and high-cost tools; when limits are exceeded, fail closed with HTTP 429 and Retry-After and do not execute the tool.

What to check

• Rate-limit middleware or equivalent is present on discovery and invocation endpoints in server code, not just in ingress or proxy config.
• Limits are keyed by identity and session, with tighter budgets for write/high-cost operations.
• Exceeded requests stop before backend action and return 429 with Retry-After.

Starting thresholds (tune to actual load, downstream limits, and cost):

Key pitfall: Gateway-only throttling or one flat bucket leaves bypasses and under-protects expensive tools.

MCP-04 — Schema validation

Scope: MCP servers exposing tools with structured arguments

Condition

• Validate all tool arguments against explicit schemas before execution.
• Schemas define types, required fields, enums, and bounds, and reject unspecified properties by default (additionalProperties: false or equivalent).
• Validation runs server-side on every invocation; invalid input fails closed with a 400/MCP error and no backend action.

What to check

• Each tool descriptor has a schema covering types, required fields, enums, bounds, and property restrictions.
• Validation occurs at the server boundary on every call, not only in clients, gateways, or downstream services.
• Negative tests reject malformed input, extra properties, and bounds violations.

Key pitfall: Allowing extra properties or client-only validation creates hidden attack surface and scope creep.

MCP-05 — SDK-first

Scope: Remote MCP servers

Condition

• Build remote MCP servers on an official MCP SDK for your server's language:
  • Tier 1 (fully supported): TypeScript (modelcontextprotocol/typescript-sdk), Python (modelcontextprotocol/python-sdk), C#/.NET (modelcontextprotocol/csharp-sdk), Go (modelcontextprotocol/go-sdk)
  • Tier 2/3 (developing): Java (modelcontextprotocol/java-sdk), Kotlin (modelcontextprotocol/kotlin-sdk), Rust (modelcontextprotocol/rust-sdk), Swift (modelcontextprotocol/swift-sdk), PHP (modelcontextprotocol/php-sdk), Ruby (modelcontextprotocol/ruby-sdk)
• If not using an official SDK, mark MCP-05 as NEEDS INVESTIGATION.
• Keep the SDK current and patched, and verify which controls are automatic versus manual.

What to check

• Dependencies reference an official MCP SDK rather than a hand-rolled HTTP/SSE stack.
• If no SDK is used, the repo contains direct evidence for auth/authz, sessions, rate limits, and schema validation.
• Dependency pinning and update hygiene show the SDK is maintained.

Key pitfall: Hand-rolled servers often miss one "small" primitive—per-request auth, throttling, or validation—and the gaps compound.