Vulnerability  ·  2026-08-19

Onyx AI platform MCP endpoints leak other users' OAuth Authorization headers

VulnerabilityHigh impactGlobalCVE-2026-71424
CVE-2026-71424 (CVSS 9.6 Critical), published 2026-08-17, is a cross-user credential leak in Onyx's Model Context Protocol integration, where per-user OAuth tokens for MCP servers are stored in and served from a shared configuration object, letting a low-privileged user impersonate another user's authenticated connection to external MCP tool servers.
MCP servers are the standard mechanism for connecting AI agents to external tools and data sources; leaking another user's OAuth token lets an attacker impersonate that user against every connected external system (e.g. SaaS tools, internal APIs) reachable through Onyx's agent tooling.
GET /api/mcp/servers and GET /api/mcp/servers/persona/{persona_id} endpoints return a shared admin MCPConnectionConfig row's auth_template.headers because OnyxTokenStorage.set_tokens/set_client_info incorrectly copy per-user OAuth tokens into a shared config row; any BASIC_ACCESS user querying these endpoints receives another user's OAuth Authorization header for connected MCP servers.
Onyx (open-source AI platform), prior to 3.1.10, 3.2.14, and 4.0.0
Upgrade to Onyx >= 3.1.10 / 3.2.14 / 4.0.0; rotate OAuth tokens for connected MCP servers and audit MCP server access logs.
TheHackerWire - Onyx AI Platform Critical OAuth Token ExposureCVE Record
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →