ai-safe2-framework

CP.5.MCP: MCP Server Security Profile

AI SAFE² version: 3.1 MCP specification: 2026-07-28 primary; 2025-11-25 legacy compatibility Control family: CP.5 Server and Tool Security Profile controls: MCP-1 through MCP-19

Normative scope

This profile governs the agent-to-tool enforcement plane for Model Context Protocol deployments. AI SAFE² controls define required outcomes and evidence. NEXUS is CSI’s reference implementation; organizations may use NEXUS or another implementation that demonstrably satisfies the applicable controls.

Protocol-independence rule

A CP.5 profile MUST NOT bind a control to a construct owned by the protocol it profiles. Governance therefore binds to verified principals, capability grants, provenance baselines, delegation chains, state handles, and explicit trust-establishment events rather than transport sessions or vendor lifecycle constructs.

MCP 2026-07-28 removes the required initialize/initialized handshake and Mcp-Session-Id. Optional protocol features such as server/discover MUST NOT become mandatory AI SAFE² presence requirements.

Controls

ID Control Required outcome
MCP-1 Command Construction Safety Tool servers do not construct executable commands from untrusted arguments without strict allowlisting and escaping.
MCP-2 Return-Path Content Sanitization Tool output is treated as untrusted content and inspected before entering model context.
MCP-3 Least-Privilege Tool Exposure A principal can reach only explicitly granted tools and resources.
MCP-4 Server/Binary Integrity Local and stdio MCP servers are verified before execution; endpoint identity is verified for remote transports.
MCP-5 Tool Invocation Audit Every protected invocation emits attributable, reconstructable evidence.
MCP-6 Input Validation Tool inputs are validated against the authorized schema and policy before dispatch.
MCP-7 Trust Establishment A verified principal and policy context exist before protected tool use.
MCP-8 Economic Ceiling Consumption is accounted to a principal and fails closed at the authorized ceiling.
MCP-9 Secret Boundary Secrets are not exposed through arguments, logs, model context, or untrusted tool output.
MCP-10 Delegation Edge Monitoring Agent-to-tool actions retain the originating delegation lineage.
MCP-11 Catalog Provenance Tool/resource/prompt catalogs are compared to an authorized provenance baseline at trust establishment and revalidation.
MCP-12 Principal-Scoped State Durable or handle-scoped state is bound to the verified principal and policy context.
MCP-13 Revocation and Reauthorization Revocation or material policy change invalidates prior authorization without relying on a transport session boundary.
MCP-14 Extension Capability Negotiation Extensions are reachable only when included in the principal’s capability grant and recorded at trust establishment.
MCP-15 Header/Body Assertion Integrity Mcp-Method and Mcp-Name, when present, agree with the JSON-RPC body; disagreement is rejected before dispatch.
MCP-16 State Handle Binding and Lifecycle State handles are unguessable, expiring, principal-bound credentials and are never accepted as identity by possession alone.
MCP-17 MRTR Round-Trip Integrity Model-mediated tool responses are bound to the originating request, validated as untrusted input, and protected against replay.
MCP-18 Catalog Cache Integrity Cache TTLs are policy-bounded; cache keys include trust context; revalidation detects catalog/schema provenance drift.
MCP-19 Authorization Chain Integrity Protected requests validate issuer, intended resource/audience, redirect/resource binding, authorization metadata, and SSRF boundaries before state change.

MCP-19 conformance note

Opaque static bearer tokens do not contain an audience claim. Deployments using such tokens MUST NOT claim MCP-19 audience-validation conformance unless equivalent resource binding is independently established and evidenced. The AI SAFE² reference server’s legacy static-token mode is therefore compatibility-only for MCP-19. JWT/OAuth deployments must validate the intended audience/resource before dispatch.

Evidence minimum

Evidence for this profile SHOULD identify framework_version, control_profile_version, mcp_spec_version, principal_id, capability_grant_id, delegation_chain_id, provenance_baseline_id, trust_establishment_id, mcp_transport, mcp_server_identity, catalog hashes, cache revalidation results, OAuth resource/audience results when applicable, and policy/adapter/grader hashes.

Compatibility window

The 2025-11-25 transport binding remains supported for twelve months from the v3.1 release. Mcp-Session-Id in that binding is treated only as a principal-scoped state handle. It is not the governance identity or authorization boundary.

Reference implementation

CSI’s NEXUS MCP adapter is the first-party reference implementation path for the agent-to-tool plane. The v3.1 adapter contract is fail-closed and currently marked implementation scaffolding. Unimplemented methods raise rather than silently allowing traffic. Production deployments must use an implemented and tested enforcement component.

AI SAFE² v3.1, Cyber Strategy Institute