Skip to content

Security: nestm-dev/mcp

Security

SECURITY.md

Security policy

NestM MCP sits between agents and executable capabilities. Treat its configuration, credentials, server registry, and policy hooks as security-sensitive infrastructure.

Supported versions

During the alpha, only the latest 0.1.0-alpha.* release receives security fixes. Older alpha builds and source snapshots are unsupported. This policy will be revised before a stable release.

Report a vulnerability

Use GitHub private vulnerability reporting. Do not open a public issue, discussion, or pull request containing exploit details, credentials, access tokens, private server URLs, or customer data.

Include, when possible:

  • affected package and version;
  • deployment shape and transport;
  • minimal reproduction or proof of concept;
  • expected and observed authorization boundary;
  • impact and required attacker access; and
  • any temporary mitigation already applied.

Maintainers will acknowledge the report through the private advisory, investigate across every affected package, coordinate a fix and disclosure, and credit the reporter if requested. Please allow time for a coordinated release before publishing details.

Security boundaries

  • OAuth authentication does not replace application authorization. A valid bearer token identifies a caller; a policy must still authorize that caller for the requested server, capability, resource, or tool.
  • An MCP session identifier is routing metadata, not an authentication credential.
  • A gateway plays both roles: it is a resource server to downstream callers and an OAuth client to upstream servers. Incoming bearer tokens must not be forwarded upstream by default.
  • Stdio transport can execute a local process. Server command, arguments, working directory, and environment must come from trusted configuration.
  • An HTTP MCP server URL is an outbound network destination. Production registries should use explicit allowlists and protect against DNS rebinding and private-network SSRF.
  • Tool descriptions and tool results are untrusted content. They can contain prompt injection even when the server itself is authenticated.

Operational guidance is maintained in Security and OAuth.

There aren't any published security advisories