Commercial verification engagements — USD 8,000–15,000, court-ready dossier · Independent audit & conformance
Naming the "Session-as-Identity" antinomy — and giving every MCP server a free checkup.
The claim: MCP conflates a session routing identifier (
Mcp-Session-Id) with a security subject identifier. This architectural antinomy — Session-as-Identity — explains an entire defect class across SDKs and deployments. We provide a cross-implementation test protocol, a replayable audit suite, and a remediation matrix.
Formal definition (Phase 1 lock): DEFINITION.md
Counterexample register: PHASE1-COUNTEREXAMPLES.md
90-day invention roadmap: INVENTION-ROADMAP.md
curl -X POST https://api.correctover.com/v1/audit \
-H "Content-Type: application/json" \
-d '{"url": "https://your-server.example/mcp"}'Three verdicts: unauthenticated-surface / stateless-no-session / bound-or-inconclusive. No signup, no key. Full API · What a paid dossier looks like (free sample)
Track record (verifiable, external): participating contributor in the decision-ledger RFC discussion at google/adk-python #6099 (in-thread since Aug 2026); 4th independent author-set in the action-ref conformance ecosystem (argentum-core #163, draft-etcheverry-action-ref #12); IETF draft-correctover-ccs rev 10 (expires 2027-04).
- The session ID answers "which context am I operating in"; authentication answers "who is calling". The MCP Streamable HTTP specification (2025-06-18) never requires these to be bound, and the official TypeScript SDK validates a session with plain string equality only.
- The consequence is not theoretical: in an authenticated deployment, an attacker with their own valid token plus a leaked victim session ID can (1) execute tools inside the victim's session and (2) terminate the victim's session — both verified against the official reference deployment with full control-group evidence.
- The ecosystem already acknowledges the class (e.g. CVE-2026-67431, Ruby SDK "Session Poisoning"; SEP-2567 removes sessions entirely and is merged as Final). What's missing is a systematic, per-implementation, testable matrix — that is this project.
| Path | What it is |
|---|---|
DEFINITION.md |
Formal definition of the Session-as-Identity antinomy + Bound criteria |
PROTOCOL.md |
Standardized experiment protocol (P-A0 … P-LEID) — run it against your server |
PHASE1-COUNTEREXAMPLES.md |
Counterexample register and control-group checklist |
INVENTION-ROADMAP.md |
90-day path toward Subject-Bound Session Receipt |
tool/ |
Audit CLI: point it at an MCP endpoint, get a matrix row back |
MATRIX.md |
Remediation matrix: SDKs × deployment modes × binding/removal strategies |
THREAT-MODEL.md |
Precondition analysis: where session IDs leak, what an attacker needs |
evidence/ |
Audit dossiers (self-operated targets; replayable) |
- Uses
Mcp-Session-Idat all? (stateless / sessionless deployments) - Binds session to authenticated subject? (subject-swap test)
DELETE /mcpre-authenticates?- Control group: no-token request → 401? (proves auth layer is real)
- Tool execution attribution: by session owner or by token? (privilege-escalation vs. integrity-only)
- Remediation strategy: subject binding or session removal (SEP-2567, merged Final)?
- Formal definition locked (DEFINITION.md)
- Reference-implementation dynamic study (TypeScript SDK class, multiple modes)
- Authenticated-subject-swap control-group experiment
- Attribution qualification experiment
- CX-3 second auth-enforced counterexample (Phase 1 open item)
- Cross-implementation matrix v1 (target: 10+ implementations)
- Subject-Bound Session Receipt design (Phase 2)
- Release of audit CLI v0.1
- Responsible-disclosure wave + SEP input
Need your MCP server or deployment audited against this protocol? Correctover provides manual security audits of agent/MCP integrations — fail-closed review with an evidence dossier you can re-verify. Rates on request: wangguigui@correctover.com.
Backup endpoint: https://correctover-api.pages.dev/v1/audit. Methodology in this repo.
Guigui Wang — Correctover · wangguigui@correctover.com
License: CC BY 4.0 (docs) / MIT (code). All targets in published studies were self-operated reference deployments unless noted; no third-party systems tested without authorization. Not first discovery of the pattern; contribution is protocol, matrix, and binding path.
DEFINITION.md— formal definition + verdict labels (Phase 1 locked)INVENTION-ROADMAP.md— 90-day plan, pre-registered success criteria (reconciliation 2027-02-14), evidence-tier disciplinePHASE1-COUNTEREXAMPLES.md— CX register; Phase 1 CLOSED 2026-11-16 (CX-1 TS + CX-3 Python, both auth-enforced)SBSR.md— Phase 2 design v0.2 + TV-B1..B5 falsifiable vectors + dual-verifier gateevidence/cx1-ts-sdk/— TS SDK 1.32.2 dynamic verdict (measured dynamic)evidence/cx3-python-sdk/— Python SDK 2.3.0 dynamic pack (measured dynamic)evidence/sbsr/— fixtures (17) + dual verifiers + battery-results.json (zero disagreements)
Terminology note: "ACM CCS" always refers to the ACM Conference on Computer and Communications Security. The project abbreviation "CCS" (Correctover Conformance Suite, full name TBD by owner) is a separate term; docs avoid conflating them. SBSR full name: Subject-Bound Session Receipt.