CVE-2026-115563: OAuth credential leak in Model Context Protocol Python SDK

OAuth client support in the official Model Context Protocol Python SDK could send client secrets, authorization codes or signed assertions to an authorization server picked by an untrusted MCP server. Fixed in 1.30.0 and 2.2.0.

Christopher Ort
By Christopher Ort · 2 min read
Share
A developer reviewing security alerts on a laptop screen in a dimly lit office
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. What you should do
  5. What it means
  6. Sources
i10X Quick Take

Quick Take

  1. What: CVE-2026-115563, an OAuth credential disclosure flaw in the Model Context Protocol Python SDK.
  2. Who is affected: HTTP clients using ClientCredentialsOAuthProvider, PrivateKeyJWTOAuthProvider, OAuthClientProvider, or the deprecated RFC7523OAuthClientProvider from version 1.9.1 until the fixed releases 1.30.0 and 2.2.0.
  3. What to do: Upgrade to 1.30.0 (1.x) or 2.2.0 (2.x) now. Pass the issuer= parameter when constructing OAuth providers to bind credentials to a specific authorization server.
  4. Not affected: MCP servers, stdio clients, and clients that attach their own tokens or headers are not vulnerable.
  5. Severity: High (CVSS 7.5).

In plain English

A security flaw in the Model Context Protocol Python SDK allowed OAuth credentials (client secrets, authorization codes, or signed assertions) to be sent to the wrong authorization server. This happened because the SDK did not properly check which server should receive the credentials, so an untrusted MCP server could trick the client into sending its credentials to an attacker-controlled server.

What you should do: If you use the MCP Python SDK with OAuth for HTTP clients, you should upgrade to the fixed version right away. If you only use MCP servers or stdio clients, you are not affected and can ignore this issue.

What happened

The Model Context Protocol Python SDK from version 1.9.1 up to the fixed releases 1.30.0 (1.x line) and 2.2.0 (2.x line) contained a vulnerability in its OAuth client support. The affected OAuth providers (ClientCredentialsOAuthProvider, PrivateKeyJWTOAuthProvider, OAuthClientProvider, and the deprecated RFC7523OAuthClientProvider) could not properly identify which authorization server owned their credentials.

As a result, an untrusted MCP server could direct the client to send sensitive OAuth credentials to an authorization server of the attacker’s choosing. The credentials at risk included client secrets, authorization codes, PKCE code verifiers, and signed client assertions.

The vulnerability was addressed in pull requests #3398, #3431, and #3435. The fixes validate the authorization server metadata issuer on every discovery path, ensuring that credentials are only sent to the correct authorization server. The SDK now checks the metadata’s issuer field against the expected authorization server URL, as required by RFC 8414 section 3.3.

What you should do

Upgrade to Model Context Protocol Python SDK version 1.30.0 or 2.2.0. When constructing ClientCredentialsOAuthProvider or PrivateKeyJWTOAuthProvider, pass the issuer= parameter naming the authorization server that issued your credentials. This binds the credentials to that specific server and prevents them from being sent elsewhere.

Constructing these providers without the issuer= parameter is now deprecated and will be required in version 3.0.

What it means

This vulnerability allowed an attacker controlling an MCP server to steal OAuth credentials from clients. The impact is high because OAuth credentials grant access to protected resources and can be reused by attackers. Organizations using the affected SDK versions for HTTP-based MCP clients should audit their deployments and rotate any credentials that may have been exposed.

The fix ensures that OAuth clients only send credentials to the authorization server that issued them, preventing credential theft through malicious or compromised MCP servers.

Malicious MCP servers stealing OAuth tokens from clients is a pattern across SDKs: see the token theft flaw in the mcp-go OAuth client and the MCP TypeScript SDK delivering messages to the wrong HTTP client. New cases are tracked in our AI Security hub.

Sources

NVD entry CVE-2026-115563; modelcontextprotocol/python-sdk release notes v1.30.0 and v2.2.0; python-sdk pull requests #3398, #3431 and #3435.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. CVE-2026-126681: Prompt injection flaw in AWS Labs Timestream for InfluxDB MCP Server leaks API tokens
  2. CVE-2026-115563: OAuth credential leak in Model Context Protocol Python SDK
  3. US and Nine Allies Sign San Francisco Declaration to Protect AI Supply Chains
  4. Apple Launches Health AI Study With OpenEvidence for Up to 200,000 iPhone Users
  5. AMD Completes $8.2 Billion World Labs Acquisition, Fei-Fei Li Becomes Chief Scientist