// Security Advisory Triage

MCP SDK OAuth flaw lets servers redirect credentials to attacker-controlled endpoints

GitHub advisories — npm · published 2026-10-07 06:04 UTC
ACTION NEEDED PATCH_NOW SECURITYAPI
WHAT CHANGEDGitHub published CVE-2026-104850 on 6 October 2026, rating it high severity. The vulnerability affects @modelcontextprotocol/sdk and @modelcontextprotocol/client: an MCP server can specify an authorization server URL, causing the OAuth client to send credentials to an attacker-controlled endpoint. Patched versions are 1.31.0 and 2.2.0.
WHY IT MATTERSAny application using the MCP TypeScript SDK for OAuth flows can leak credentials to a malicious or compromised MCP server. The server controls where the client sends tokens, bypassing the intended authorization boundary. This is a credential exfiltration path in prompt/tool/credential handling workflows.
WHAT TO DOGrep your package.json and lock files for @modelcontextprotocol/sdk or @modelcontextprotocol/client. If present and below 1.31.0 or 2.2.0 respectively, update to the patched version and redeploy. If you cannot update immediately, audit which MCP servers your application connects to and remove any you do not control.
Model-reported confidence 95% (not measured accuracy) Primary source ↗ written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
+ghsa_id: GHSA-6qxp-vccf-f47h
+summary: MCP TypeScript SDK: OAuth client could send credentials to an authorization server chosen by the MCP server
+severity: high
+published_at: 2026-10-06T15:35:44Z
+cve_id: CVE-2026-104850
+vulnerabilities[].package.name: @modelcontextprotocol/sdk, @modelcontextprotocol/client
+vulnerabilities[].first_patched_version: 1.31.0, 2.2.0
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
triage verdictMATERIAL at 0.92, needed 0.5
triage reasoningMCP SDK vulnerability in OAuth credential handling affects AI agent frameworks and integrations that security leads deploy.
signalsAI provider SDK affected (@modelcontextprotocol packages) · OAuth credential handling vulnerability · High severity rating · Fixed versions available (1.31.0, 2.2.0) · New advisory with CVE assignment · Affects prompt/tool/credential handling in MCP framework
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)