Skip to content

Security: stellar-experimental/stellar-raven

SECURITY.md

Security policy

This repository runs a live public service: the Stellar Raven MCP gateway at raven.stellar.org. The gateway is a Cloudflare Worker. It is an OAuth authorization server, and it runs model-written code in a sandbox. We welcome security reports.

Report a vulnerability

Do not open a public issue for a security problem.

  • Use GitHub private vulnerability reporting on this repository (Security tab, then "Report a vulnerability"). This is the preferred channel.
  • Or send an email to frontier@stellar.org. Include a description, the steps to reproduce, and the impact that you expect.

We acknowledge a report within a few business days. Give us a reasonable time to fix the problem before you disclose it publicly.

For other problems, such as connection problems, catalog questions, or general support, use the #raven channel in the Stellar Developers Discord. Do not post a vulnerability there.

Scope

In scope:

  • The source code and generated artifacts in this repository.
  • The gateway at raven.stellar.org, and any retired hostname that still routes to it. This includes the authentication flows, the search and execute MCP tools, the /playground page with its login and chat routes, and sandbox isolation and egress.

Out of scope:

  • The upstream services that the gateway uses: Lumenloop, Stellar Light/Scout, and Stellar Docs. Report those problems to their owners.
  • Attacks that need a compromised maintainer machine.
  • Volumetric denial of service.

Notes for researchers

  • Model-written code runs in a Dynamic Worker isolate with no network egress (globalOutbound: null). Host-side adapters make all service calls and hold the secrets. Sandbox-escape and egress findings are the most valuable reports.
  • ARCHITECTURE.md documents the authentication design: WorkOS OAuth, named API keys, and a development bypass. The bypass works only on loopback hosts. The design does not depend on secrecy. A report that only restates documented behavior is not a vulnerability. An example is a report that named API keys or the loopback-only bypass exist.

There aren't any published security advisories