MCP Transport Hardening

Control objective

MCP Transport Hardening checks that every connection over the Model Context Protocol — the channel an AI agent uses to reach tools, document-extraction services and RAG knowledge stores — is encrypted, authenticated and tamper-evident, so an attacker cannot eavesdrop, impersonate a server, or inject malicious tool responses. This matters because in BFSI an unprotected MCP session is a direct path for adversary-in-the-middle attacks (the risk pattern behind OWASP ASI06 and MITRE ATLAS T0050): a forged tool reply could feed a loan-summarisation agent fabricated account data. The control is binary rather than statistical: the target is that the MCP transport is hardened with TLS, mutual authentication and integrity protection, and that there are 0 unauthenticated transport sessions. To implement, terminate all MCP traffic over TLS, require client/server credentials (mTLS or signed tokens) before any session is established, and add message-integrity checks (signatures or HMAC) on payloads; instrument the gateway to log every session's auth status, cipher and integrity-verification result. Operate it by continuously counting sessions that lack authentication or encryption. Because the threshold is zero, any single unauthenticated or unencrypted session is a breach that must immediately drop the connection, raise a security alert and quarantine the affected tool route pending investigation.