Security model
Trust boundaries, fail-closed defaults, what the ATP MCP server refuses, and the known limitations you should plan around.
This page explains how the ATP MCP server limits what an AI agent can do, and where those limits stop. Read Authorization and Adapter API first if you have not.
Not deployed
Maturity
The controls below are implemented and tested locally against in-memory fakes only: a fake chain RPC, a fake application backend and locally generated tokens. The server is not deployed. Execution is gated off by default. No application implements the adapter API yet. No independent security review has been done.
Trust boundaries
The server treats the model, the MCP host and every tool argument as untrusted. Each hop has its own independent check. Past the backend, the signer and the ATP contracts re-check everything themselves.
[untrusted model + MCP host]
| HTTPS, bearer token
v
[1 ATP MCP server] -- read-only RPC --> [3 chain: finalized block, pinned runtime]
| per-application service credential, fixed paths
v
[2 owning application backend] -- its own private path --> [4 signer] --> [5 ATP contracts]| Boundary | Untrusted input | Check on the receiving side |
|---|---|---|
| Host to server | Prompts, tool arguments, token, headers | Host, Origin, size and type guards, token validation, client registry, exact scope, operator gate, strict schemas |
| Server to backend | Job, session and request references | The backend maps the caller to an account and checks ownership. The server validates every response and cross-checks the echoed reference |
| Server to chain | RPC answers | Chain id, contract code hash, protocol version, configuration id, strict ABI decoding, finalized block pinned |
The server never holds a private key, never calls a signer, never forwards the caller's bearer token, never writes chain state, and stores no job or idempotency state. It accepts no amount, payee, wallet, token, calldata, signature or URL in any tool. The application comes only from the server-side client registry, never from tool input. See Tools for the six tools.
Fail-closed defaults
When something is missing or unclear, the server denies.
- A configuration problem stops the process. An enabled gate without a backend stops the process. An empty client registry denies every token.
- Any guard failure happens before the SDK and makes zero chain calls and zero backend calls.
- Policy is recomputed on every call. An unknown tool, unknown application, missing exact scope, closed gate, failing rate limiter or failing verifier all deny.
- Backend and chain answers are validated strictly. An unknown field, a different echoed reference, a binding for another chain or an evidence block from another deployment fails the call.
- A chain read that cannot prove finality, freshness, chain id or runtime match fails. There is no weaker best-effort result.
- Errors carry only a category, a stable reason code, a retry flag and a correlation id. See Schemas and contracts.
- On the execute path, any outcome that is not a provable non-send becomes
RECONCILIATION_REQUIREDand is never retry-safe. The server never retries.
What the server refuses
The request pipeline runs in a fixed order. A denial at any step stops the request.
| Step | Rule | Answer |
|---|---|---|
| Host | Must be exactly host[:port] and on the allow-list | 403 |
| Origin | If present, must match an allow-listed origin exactly. null is rejected | 403 |
| Content type and body size | Must be JSON, and counted byte by byte against a limit | 400, 413, 415 |
| Bearer | Exactly one Authorization: Bearer header. No token in the query string | 401 with challenge |
| Token | Signature, issuer, audience equal to this resource, expiry, client id, scopes, and a client registry lookup | 401, 403, 503 |
| Rate limit | Charged per verified issuer, subject and client | 429 |
| Tool policy | Known tool, known application, exact scope, operator gate for preflight and execute | Error result |
Cross-tenant access is refused in layers. The server picks the backend from the registry, so a caller cannot reach another application's backend. A reference that belongs to another account, or does not exist, returns the same REFERENCE_NOT_ACCESSIBLE answer with no detail. Whether a backend truly enforces tenant isolation is the backend's job, and no backend has proven it yet.
Redirects are refused, and backend URLs use only five fixed paths on the configured origin.
Audit and redaction
Every call produces an audit event with allow-listed fields only. Secrets, tokens, lease tokens and signatures are redacted from errors, outputs and logs, and tests scan every response and audit line for leaks. A failing audit sink or redaction step never changes an allow or deny decision. An unexpected exception becomes a generic 500 with only the correlation id.
Known limitations
Plan for these before you connect an application.
| Limitation | What it means |
|---|---|
| Per-process rate limiter | Buckets reset on restart and are not shared across replicas. Use a proxy for real quotas. It is not an idempotency mechanism |
| No live-chain evidence | The pinned deployment and ABI were derived offline. The reader ran only against a fake RPC. A first live read is still required |
| Adapter API unimplemented | Most tools answer BLOCKED_INTEGRATION until an application builds the adapter API |
| No DNS-level egress pinning | The code uses configured https origins but does not reject private IP ranges. Enforce an egress allow-list in the network |
| No independent security review | All verification is the authors' own tests |
| Token revocation not consulted | Tokens are valid until expiry, at most 24 hours. Remove the client from the registry to cut it off |
| Finality lag | Reads use the finalized block, so a new Atom or a fresh revocation may not be visible yet |
Negative-test categories
The test suite covers attacks and faults by category. Proven here means a local test passes against fakes. Partly means the server's side is proven and another owner's side is not.
| Category | Scenario | Status |
|---|---|---|
| Protocol | Pre-2026 handshake, oversized body, wrong content type, bad Origin or Host | Proven here, except missing per-request metadata (left to the SDK) |
| Input | Unknown tool, extra fields, bad Atom locator, poisoned text | Proven here |
| Authentication | Missing, expired, forged or wrong-audience token | Proven here |
| Authorization | Read-scoped token calling execute | Proven here |
| Tenancy | Cross-tenant job, wrong application, forged account fields | Partly. A backend must prove the denial |
| Execution | Concurrent requests, timeout after send, ambiguous outcome, reorg | Partly. One request, no retry, RECONCILIATION_REQUIRED on doubt. No real reorg was tested |
| Egress | Redirects, foreign origin, SSRF, DNS rebinding | Partly. Fixed paths proven, IP and DNS controls not in code |
| Leakage | Secrets in errors, outputs or logs | Proven here |
| Operations | Execution gate closed | Proven here |
Controls that belong to a backend, the signer or the contracts are outside this server. Re-run the suite for current counts.