Security incidents involving Model Context Protocol servers and tooling: supply-chain campaigns, credential disclosure, path traversal, and SSRF. Each entry links to primary sources and the disclosed CVE where one exists.
HTTP handlers can accept bodies or JSON-RPC batches without sufficient bounds. A reachable client can use oversized input to consume memory or processing capacity before normal tool handling.
A request can satisfy the default byte limit yet cause excessive memory use while the SDK validates it. Size checks alone do not establish a safe bound on the cost of processing an untrusted request.
Reusing an HTTP server or transport across clients allows overlapping request identifiers to associate a response with the wrong connection. Applications must isolate server and transport state per request or session.
A token issued for a different service can be accepted when a shared authorization server is used and the application does not validate the intended resource. Upgrading provides the control, but the advisory says it must also be enabled and configured.
Server authentication can accept a valid token without establishing that the token was issued for this MCP resource. Deployments using a shared authorization server need an explicit audience check as well as the maintained SDK versions.
Loom for AWS exposed a fixed super-admin identity when no Amazon Cognito user pool or external identity provider was configured. A network client could then read, create, modify, or delete platform resources without credentials.
The HTTP transports and OAuth request paths can follow redirects that resend custom headers or method-preserving request bodies to a different origin. A redirect under attacker influence can expose that material.
Clients sharing an experimental task store can reach tasks created in another session. This crosses the expected ownership boundary for listing, reading or cancelling work.
Cross-origin redirects can carry application-added headers and, for redirects that preserve the method, request bodies to another host. This can expose custom API keys and OAuth exchange material; the ordinary Authorization header is a separate case.
The server reads an HTTP request body into memory before enforcing an adequate limit. A large request can repeatedly exhaust memory, including on reachable OAuth endpoints that do not require credentials.