CVE-2026-53512: Better Auth: OAuth refresh-token replay via missing client authentication on oidc-provider and mcp plugins

Am I affected?

Users are affected if all of the following are true:

- Their application uses better-auth and has enabled at least one of: oidcProvider() (imported from better-auth/plugins/oidc-provider), or mcp() (imported from better-auth/plugins/mcp). - Their application has at least one confidential OAuth client registered (any client with type: "web" | "native" | "user-agent-based" in the oauthApplication table, or any trustedClients entry without type: "public"). Public clients with PKCE are not affected. - Their application uses better-auth at a version below the patched release.

If an application only uses @better-auth/oauth-provider (the canonical replacement for oidc-provider) and the mcp plugin is not enabled, it is not affected.

Fix:

1. Upgrade to better-auth@1.6.11 or later. 2. Migrate from the deprecated oidcProvider() to @better-auth/oauth-provider when feasible. The new package enforces client authentication on both grants by default. 3. If developers cannot upgrade their applications, see workarounds below.

Summary

The legacy oidcProvider and mcp plugins each expose an OAuth 2.0 token endpoint whose refreshtoken grant authenticates the request entirely on possession of the bound refreshToken row and a matching clientid. Neither plugin verifies the registered confidential client's clientsecret on the refresh path. An attacker who obtains any valid refreshtoken (via database read, log capture, browser-side XSS, or CORS-amplified script in the mcp case) and the public clientid can mint fresh access tokens and rotated refresh tokens until the chain is revoked.

Details

RFC 6749 §6 and OAuth 2.1 §4.3 require confidential clients to authenticate to the token endpoint on every grant, including refresh. The same plugins' authorizationcode grant correctly enforces clientsecret (the oidc-provider via verifyStoredClientSecret, the mcp plugin via raw equality), which proves the omission on the refresh path is a regression rather than a design choice.

Token rotation issues a new refreshtoken with each call, so a single leaked refresh-token grants indefinite access until the row is revoked or its refreshTokenExpiresAt (default 7 days) passes; rotation refreshes that window each call.

Two adjacent issues on the mcp surface ship in the same patch. The mcp authorizationcode grant uses raw === for client-secret comparison and ignores the storeClientSecret: "encrypted" | "hashed" configuration; the fix routes both grants through verifyStoredClientSecret. The mcp /mcp/token endpoint sets Access-Control-Allow-Origin: unconditionally, which amplifies the refresh bypass in browser contexts; the fix narrows the CORS allowlist.

The newer @better-auth/oauth-provider package routes both grants through validateClientCredentials and is not affected.

Patches

Fixed in better-auth@1.6.11. The legacy oidcProvider and mcp token endpoints now require clientsecret on the refreshtoken grant for confidential clients, using the same constant-time comparison the authorizationcode grant already used. Public clients are unaffected (they have no secret to enforce, and PKCE substitutes on the auth-code grant).

The Authorization: Basic parser is fixed to follow RFC 6749 §2.3.1: the credential is split on the first colon and each half is percent-decoded. Client IDs and secrets that contain reserved characters now authenticate correctly. The /mcp/token endpoint's CORS configuration is narrowed in the same change (the wildcard Access-Control-Allow-Origin: header is removed), matching the standalone @better-auth/oauth-provider package.

The deprecated oidc-provider plugin remains deprecated. The recommended migration path is @better-auth/oauth-provider.

Workarounds

None of these close the bug fully without a code patch.

- Migrate to @better-auth/oauth-provider if your deployment can adopt the new plugin. It enforces clientsecret on both grants. - Force all clients to public + PKCE: set every client's type: "public" and require PKCE. The bug is unreachable when there is no clientsecret to verify. - Network-layer ingress restriction: limit /api/auth/oauth2/token and /api/auth/mcp/token to known client IPs at the load balancer. Practical for server-to-server flows, not for end-user-device clients. - Out-of-band refresh-token rotation: on any suspicion of leak, run db.deleteMany({ model: "oauthAccessToken", where: [{ field: "clientId", value: <id> }] }) to invalidate all refresh tokens for the affected client. - For the mcp endpoint specifically: drop the wildcard CORS at an upstream proxy and replace with a tight allowlist.

Impact

- Indefinite confidential-client impersonation: an attacker holding any valid refreshtoken and the public clientid can mint access tokens and rotated refresh tokens indefinitely, until the row is revoked. Rotation refreshes the expiration window each call. - Resource access at the user's authorized scope: every minted access token carries the original user's authorization scope, so the attacker reads or writes whatever the resource server grants for that scope.

Credit

Reported by @subhanUmer.

Resources

- CWE-306: Missing Authentication for Critical Function - CWE-287: Improper Authentication - CWE-345: Insufficient Verification of Data Authenticity - CWE-863: Incorrect Authorization - RFC 6749 §6: Refreshing an Access Token - OAuth 2.1 §4.3: Refresh Token