TL;DR — We measured session/subject decoupling in both official MCP SDKs on auth-enforced deployments (cross-subject tool execution and session termination), built a remediation object (SBSR) with falsifiable test vectors, and ran two independent verifiers over them with zero disagreements. Everything is public and replayable: github.com/DSHCorrectover/mcp-session-audit

What we measured (self-operated targets only)

In the MCP 2025-06-18 spec, session management and authorization live in independent sections; no normative text requires the server to bind Mcp-Session-Id to the authenticated principal. Both SDKs resolve the ambiguity the same way — session-id string equality is the whole admission decision:

  • TypeScript SDK 1.32.2 — validateSession: string equality only. Measured: attacker's own valid token + victim's session id executes a tool (200), kills the session (200), the victim then gets 404. evidence/cx1-ts-sdk
  • Python SDK 2.3.0 — _validate_session (~L901): identical shape, identical results. evidence/cx3-python-sdk

Both targets are auth-enforced (401 control passes) — this is not an "auth: none" artifact.

The proposed object (SBSR)

A session-scoped decision receipt binding session_id + auth_subject + action/tool + params_hash (+ iat/exp/jti), Ed25519/JCS — existing primitives only — offline-verifiable, fail-closed on any anomaly. Design: SBSR.md

The part we care most about is the discipline: each binding claim has a mutation that must be denied (TV-B1..B5). Two independent verifiers — no shared code — pass all vectors with zero disagreements (battery results). Claim-scope note: this is an enumerated five-binding-check suite (binding / integrity / freshness), not a general proof of falsifiability — the mutation set is closed by design, and the DENY verdicts come from the implementation under test, so the self-attestation limit is real; an independent-observer deployment is the open question.. Round 1 of our own battery found that three vectors were not falsifiable as designed; we fixed the vectors and recorded the lesson rather than re-labeling it a "known limitation".

Necessity argument with a falsification clause

Why binding is load-bearing: without jointly consulting (session, subject, request), execution isolation, attribution and revocation cannot hold simultaneously. The argument — and the concrete evidence that would refute it — is written down: MINIMAL-CONDITION.md

Non-claims

No first-discovery claim (CVE-2025-53365, CVE-2026-67431, SEP-2567/1299 and related advisories are acknowledged as prior art, and are different defect classes); no new cryptographic primitive; no vendor endorsement; unauthenticated-surface defects are labeled separately.

If you maintain an MCP server implementation, the probe protocol in the repository lets you measure your own deployment in minutes. If the session-bearing direction survives, binding is the load-bearing wall.