What a Manifest Hash Verifies and What It Does Not
In MCP, a server exposes its tools through a discovery call that returns a schema-based manifest containing a tool's name, description, and JSON Schema input definition. A manifest hash is a digest computed over that data, typically using SHA-256, applied to a canonicalized form of the manifest so that field ordering or whitespace differences do not produce false mismatches.
The hash itself answers one question: does this manifest match a previously recorded state. It does not answer whether the manifest originated from a trusted server; that is a separate concern handled by transport security and server authentication. Treating hash verification as identity verification is a common implementation error, since the two mechanisms protect against different failure modes and both are required for a complete trust decision.
Attack Scenarios Addressed, and What Remains Open
Security researchers have documented two closely related patterns relevant to manifest integrity. In tool poisoning, a server presents a manifest whose description contains hidden instructions intended to manipulate agent behavior. In an MCP rug pull, a server alters a tool's description or schema after a user or client has already approved it, so the agent later acts on content that was never reviewed.
Manifest hash verification directly addresses the rug pull pattern by detecting that the approved manifest has changed, forcing re-approval rather than silent execution against altered instructions. It does not, by itself, detect tool poisoning present at initial registration, since a hash will faithfully validate a malicious manifest if that manifest was the one originally approved. Hashing is a drift-detection control, not a content-safety control, and enterprises should not treat it as a substitute for reviewing manifest content at approval time.
Implementation Decisions Enterprises Must Make
Before relying on a manifest hashing implementation, teams should be able to answer the following questions with specifics rather than assumptions:
- Is the baseline hash computed and stored independently of the MCP server providing the tool?
- At what point in the invocation path is the hash re-verified, and what is the defined response to a mismatch?
- Is the hash algorithm and canonicalization method documented and version-controlled?
- Is there a defined workflow to distinguish legitimate manifest updates from unauthorized changes?
- Are integrity controls applied consistently across both local stdio and remote HTTP/SSE server connections?