In May 2026 the United States National Security Agency published a Cybersecurity Information notice titled Model Context Protocol (MCP): Security Design Considerations
for AI-Driven Automation (document ID U/OO/6030316-26 / PP-26-1834). It is fifteen pages on what the NSA considers the minimum security baseline for any production MCPdeployment.
If you are building anything on MCP, server, client, gateway, orchestrator, framework, or agent runtime, read it. Then read this, because the standards work the NSA describes already exists, and you can integrate it today.
What the NSA called out
Four operational requirements run through the document.
- Cryptographically sign and verify MCP messages
Quoting the NSA directly (page 12): "the standard can be extended with cryptographic signatures directly within the JSON payload ... MCP messages should include expiration timestamps and replay protection metadata ... cryptographically bind requests to time and context to prevent tampering, intentional replay techniques, and
unintended re-execution."
Translation: TLS is not enough. The MCP payload itself needs an envelope with a signature, a nonce, a timestamp, and a freshness window.
- Cryptographic identity for agents (not bearer tokens)
Page 4: MCP "lacks support for exchanging Role Based Access Control permissions at instantiation." Bearer tokens can be lifted, replayed, and impersonated. Agents need
verifiable cryptographic identity, bound to scope, trust level, and issuer.
- Structured audit logging with cryptographic integrity
Page 12 to 13: log every tool invocation, every parameter, every result, with cryptographic hashes, so an XDR or SIEM can reconstruct exactly what happened and prove it
has not been altered.
- Track MCP-specific CVEs and patch them
Page 13: build a vulnerability-monitoring process around your MCP package surface, the same as you would for any other production dependency.
What already exists, today
Here is the awkward bit, depending on where you sit: every single one of these four requirements has an open specification, a reference implementation, and at least one
production integration. They predate the NSA notice.
MCPS, the cryptographic signing layer for MCP
draft-sharif-mcps-secure-mcp on the IETF Datatracker since March 2026. Four primitives:
- Agent Passports. Cryptographic identity bound to a specific origin.
- Signed message envelopes. ECDSA P-256 over a canonical-signing-string, with timestamp and nonce, for integrity and non-repudiation.
- Tool definition signatures. Covering the full tool object, so a downstream client can detect tool poisoning or schema tampering.
- Nonce plus timestamp replay protection with transcript binding, to prevent downgrade attacks.
The wire format is JSON-on-the-wire, signing-string-canonical, and stays inside the MCP message body. No transport changes, no protocol fork.
ATTP, agent-trust transport above MCPS
draft-sharif-attp, live since 1 May 2026. Where MCPS does message-level signing for MCP, ATTP defines a protocol-agnostic trust transport above it: five hierarchical
trust levels (L0 to L4), action-limit enforcement, compliance gating, and tamper-evident audit. It maps onto MCP, REST, Google A2A, gRPC, and GraphQL.
Live demo with real ECDSA P-256 in the browser, including tamper and strip-ATTP buttons:
👉 https://attp.cybersecai.co.uk
AgentPass, the identity / RBAC layer the NSA describes
L0 to L4 trust grades, OFAC and HMT sanctions screening (75,784 entries baked in), graduated spend limits, hash-chained audit trails, agent-to-agent payment
authorisation, optional Mastercard risk integration.
The Go SDK (agentpass-go) verifies agent identity certificates with zero network calls. Pure local crypto, standard library only, no CGo. Trust anchors load like TLS
root CAs.
x-agent-trust, agent trust as a first-class OpenAPI declaration
Merged into the official OpenAPI Initiative Extension Registry on 11 April 2026, approved by Henry Andrews and Mike Kistler (Microsoft):
👉 https://spec.openapis.org/registry/extension/x-agent-trust.html
components:
securitySchemes:
AgentTrust:
type: apiKey
description: Uses agent trust information in lieu of a traditional API key. Requires the x-agent-trust extension.
in: header
name: Agent-Signature
x-agent-trust:
algorithm: ES256
trustLevels: [L0, L1, L2, L3, L4]
issuerKeysUrl: /.well-known/agent-trust-keys
security:
- AgentTrust: [L3]
Any OpenAPI-described service can now declare which agent trust level is required to call which operation. Tooling that understands the extension can verify the
Agent-Signature header before the request even reaches application code.
OWASP MCP Security Cheat Sheet, Section 7
Section 7, Message-Level Integrity and Replay Protection contributed via PR #2065, merged 26 March 2026. The cheat sheet now documents the patterns the NSA later
described, including signing JSON-RPC messages with asymmetric keys, including nonces and timestamps, and pinning tool definitions using hashes.
OWASP AISVS 1.0, Chapter C10
An entire chapter on MCP Security, with verifiable requirements at L1 to L3. Two requirements map directly to the MCPS spec:
- 10.2.11 (L2). Agents authenticate using cryptographically bound identity credentials rather than bearer-only tokens, ensuring agent identity cannot be transferred, replayed, or impersonated.
- 10.4.10 (L3). MCP servers sign tool responses with a unique nonce and timestamp within a bounded time window, so the calling agent can verify origin, integrity, and freshness, preventing spoofing, tampering, and replay.
CVE-2026-39313, and five more on the clock
CVE-2026-39313. Unbounded-memory-allocation vulnerability in mcp-framework (CWE-770, High), assigned and published 16 April 2026.
Five further CVE submissions, across ~57M weekly downloads of MCP packages (including the official MCP TypeScript and Python SDKs), are under coordinated-disclosure clock.
This is the NSA's recommendation #4 in action. The package surface is being audited, and the gaps are being closed.
In production, today
moov-io / watchman (Apache 2.0, ~460 stars). Sanctions screening used by SEC-registered transfer agents and BaaS platforms. MCPS and AgentPass are merged into main. The production deployment guide ships an AgentPass configuration block:
AgentPass:
TrustAnchorPath: /etc/watchman/agentpass-ca.pem
MinTrustLevel: 2
RequiredScopes:
- sanctions:search
👉 https://github.com/moov-io/watchman/blob/master/docs/mcp.md
Cisco AI Defense. Cisco's commercial agent-security product ships our MCPS protocol as part of its agent-defence stack.
👉 https://www.cisco.com/site/us/en/products/security/ai-defense/index.html
Kong API Gateway. A plugin that turns every API behind Kong into an MCPS-signed endpoint with zero developer effort. Available to design partners under NDA.
AEBA-XDR, runtime behaviour analysis for every agent. Anomaly detection in milliseconds. Eight behavioural dimensions, every agent cryptographically identified,
hash-chained tamper-evident audit, native forwarders for major XDR and SIEM platforms via CEF, LEEF and syslog RFC 5424. Free evaluation tier for up to three agents.
👉 https://aeba.co.uk
What to do this week
If you ship MCP in production:
- Pin MCP-package versions and subscribe to CVE feeds for every MCP-related dependency. Start with CVE-2026-39313.
- Add MCPS message signing to your most-sensitive tool invocations. The spec is small enough to implement from draft-sharif-mcps-secure-mcp in a sprint.
- Declare your agent-trust requirements in your OpenAPI document using x-agent-trust. Whether or not your runtime enforces them yet, you have made the requirement machine-readable for everything downstream.
- Wire your tool invocations into a tamper-evident audit log that maps to the AISVS C10 controls.
- Read the NSA notice. If you skim only one document on MCP security this quarter, skim that one.
The standard exists. The reference code exists. The integrations exist. The CVE feed exists.
The protocol can be secured. Now there is no reason not to.
Raza Sharif (FBCS, CISSP, CSSLP)
Founder, CyberSecAI Ltd