Do not open a public issue for a security problem.
Report it privately through GitHub Security Advisories, or by email to nick@epistates.com.
Please include enough to reproduce it: the crate and version, the feature flags in play, the protocol revision if it matters, and a minimal server or client that shows the behaviour. A JSON-RPC frame or HTTP request that triggers it is worth more than a description of it.
You can expect an acknowledgement within 3 working days and an assessment within 10. If a report is accepted we will agree a disclosure date with you, publish a GitHub Security Advisory with a CVE, and credit you unless you would rather we didn't. If it is declined you will get the reasoning, not silence.
| Version | Supported |
|---|---|
4.0.x (pre-release) |
Yes — fixes land on main |
3.1.x |
Security fixes only |
< 3.1 |
No |
4.0.0-alpha/-beta/-rc releases are pre-release software. Report
vulnerabilities in them as normal, but expect fixes as a new pre-release rather
than a patch to an existing one.
In scope — anything reachable by a peer over the protocol:
- Parsing untrusted input: the JSON-RPC codec, the
Mcp-Paramheader codec, URI-template matching, and the wire-type deserializers. These are the crate's fuzz targets for exactly this reason. - Authentication and authorization: bearer-token validation, JWKS
retrieval and caching, the RFC 9728 protected-resource metadata endpoint, the
OAuth 2.1 client flow,
#[tool(scopes(…))]enforcement, and visibility policies. A component that a policy hides must be unreachable, not merely absent from a listing. - Transport handling: HTTP session management and termination,
Originand DNS-rebinding checks,X-Forwarded-Forhandling behind a trusted proxy, rate limiting, request size limits, and WebSocket framing. - Cross-session leakage: anything that lets one session observe or affect another's state, subscriptions, or MRTR turns.
- Telemetry: PII reaching spans or metrics that the redaction layer is documented to strip.
Out of scope:
- Vulnerabilities in a server you write with this SDK — a
#[tool]that shells out on unvalidated input is your bug, not the framework's. We will still take reports where the framework's API made the unsafe path the obvious one. - Denial of service through resource exhaustion that an operator can bound with the configuration the crate already provides (rate limits, size limits, session timeouts). Report it if the bound doesn't work, or if there is no bound to set.
- The
turbomcp-interopandturbomcp-conformancecrates. Both are test harnesses, excluded from the workspace, and not published. - Anything requiring an attacker who already has code execution in the server process.
Stated so you know what to hold us to, not as a guarantee:
#![forbid(unsafe_code)]in every published crate. There is nounsafeblock to audit, and adding one is a compile error rather than a review question.- Fuzz targets for every untrusted-input decoder — the JSON-RPC codec, the
Mcp-Paramheader codec, URI-template matching, and a sonic-vs-serde differential. Run out of band (just fuzz,just fuzz-long <target>) rather than on every push. cargo-denyon every push for RUSTSEC advisories, license policy, and source provenance.- A conformance suite run against the official MCP conformance harness, and cross-SDK interop tests against the official Rust SDK.
- Dependencies are chosen, not accumulated — cryptography, HTTP parsing, JSON, date/time, and IP/CIDR handling are delegated to established crates rather than reimplemented.
Several protections are opt-in because they depend on your deployment. If you
expose a server beyond localhost, review at minimum: HttpConfig's
authenticator, rate limiter, and trusted-proxy settings; the WebSocket
transport's Origin allowlist; and ServerBuilder::with_visibility if some
components should not be reachable by every caller. The crate's HTTP transport
binds where you tell it to and does not assume a reverse proxy in front.