Attack Chain

1. Initial Access / Credential Acquisition: An attacker obtains a valid refresh_token and the associated public client_id for a confidential OAuth client. This could occur through various methods such as a database breach, log file capture, browser-side Cross-Site Scripting (XSS), or via a CORS-amplified script if exploiting the mcp endpoint with its wildcard Access-Control-Allow-Origin: *.
2. Exploitation of Missing Authentication: The attacker sends a POST request to the vulnerable OAuth 2.0 token endpoint (e.g., /api/auth/oauth2/token or /api/auth/mcp/token) using the stolen refresh_token and client_id, but crucially, without providing the client_secret.
3. Token Endpoint Bypass: Due to the flaw in better-auth versions prior to 1.6.11, the oidcProvider or mcp plugins do not enforce client_secret verification on the refresh_token grant for confidential clients, allowing the unauthenticated request to proceed.
4. Access Token Minting: The vulnerable token endpoint issues a new, valid access_token to the attacker, along with a rotated refresh_token.
5. Indefinite Session Prolongation: The attacker uses the newly minted refresh_token to repeat steps 2-4, effectively resetting the expiration window of the session and gaining indefinite access.
6. Resource Access: With the valid access_token, the attacker can impersonate the legitimate client and user, making requests to resource servers to access, modify, or exfiltrate data within the scope of the original user's authorization.

Impact

The impact of CVE-2026-53512 is significant, primarily leading to indefinite confidential-client impersonation. An attacker who has acquired a single refresh_token and the corresponding client_id can continuously mint new access_tokens and rotate refresh_tokens, thereby maintaining persistent unauthorized access without further authentication. Each newly issued access_token carries the original user's authorized scope, granting the attacker the ability to read, write, or otherwise manipulate data on resource servers as if they were the legitimate user. This can result in severe data breaches, unauthorized modifications, and complete compromise of data governed by the affected OAuth client's permissions. The vulnerability affects better-auth applications leveraging the deprecated oidcProvider or mcp plugins.

Recommendation

• Upgrade immediately to better-auth@1.6.11 or later to apply the official patch for CVE-2026-53512.
• Migrate from deprecated plugins by transitioning from oidcProvider() to the canonical @better-auth/oauth-provider package, as it enforces client authentication on all grants by default.
• Implement network-layer ingress restriction for /api/auth/oauth2/token and /api/auth/mcp/token endpoints to known client IPs at the load balancer, which can partially mitigate risk for server-to-server flows.
• Force all clients to public + PKCE by setting every client's type: "public" and requiring PKCE, thereby eliminating the client_secret for verification where the bug manifests.
• For the mcp endpoint specifically, drop the wildcard CORS header at an upstream proxy and replace it with a tight allowlist to prevent CORS-amplified attacks.