Executive brief
The GitHub MCP Server, which facilitates communication between AI models and GitHub data, contains a flaw when running in multi-user HTTP mode. The system incorrectly reuses the security credentials of the very first person who logs in for all subsequent users. This can lead to incorrect security decisions, such as failing to sanitize malicious content or blocking legitimate users from accessing their own data if the first user's session expires.
Technical details
A vulnerability exists in the RepoAccessCache implementation of the GitHub MCP Server (versions 0.22.0 to 1.1.1). When running in HTTP mode with --lockdown-mode enabled, the RepoAccessCache is implemented as a process-global singleton that is initialized only once with the first authenticated user's GraphQL client. Subsequent requests from different users share this singleton, causing their lockdown-related GraphQL queries (which determine repository visibility and collaborator access) to execute using the first user's credentials. This can result in incorrect IsSafeContent decisions, potentially allowing unsanitized content from untrusted contributors to reach the AI model, or causing service denial if the initial user's token expires. The issue is fixed in version 1.1.2.
Affected products
- GitHub github-mcp-server >= 0.22.0, < 1.1.2
Timeline
- 2026-06-09: advisory: GitHub security advisory published
- 2026-06-26: disclosed: CVE-2026-48529 published to NVD
- 2026-06-26: patched: Vulnerability fixed in version 1.1.2