Skip to content

Security: Medxpert-org/lgd-engine

Security

SECURITY.md

Security Policy

Reporting a vulnerability

Please do not open a public GitHub issue for security problems.

Instead, report a vulnerability through a private GitHub Security Advisory:

https://github.com/Medxpert-org/lgd-engine/security/advisories/new

Include as much detail as you can: a description of the issue, the affected component and version, reproduction steps, and any suggested impact. Reports received through this channel are private to the maintainers until triaged.

If you cannot use GitHub Security Advisories for any reason, open a blank private issue on the repository; do not publish exploit details publicly before we have had a chance to respond.

What is in scope

Security review is requested for:

  • the judgment core (lgd_engine/gate/, lgd_engine/theory/),
  • domain judges (lgd_engine/judges/),
  • the MCP server (lgd_engine/mcp_server.py) and its smoke harness,
  • the REST/OpenAPI surface (docs/openapi.yaml, docs/validate_openapi.py),
  • the SSE event stream (demo2/sse_server.py),
  • the TypeScript SDK skeleton (lgd-engine-ts/),
  • the container definition (Dockerfile, docker-compose.yml), and
  • declared dependencies in pyproject.toml (Dependabot enabled).

Known boundary — please read before reporting

The MCP server, the REST surface, and the SSE stream are local-trust adapters. They trust the caller's role claim and bind locally (stdio / 127.0.0.1). They are not a network authentication boundary, and this is a deliberate, documented state of the 0.1 pre-release, not a hidden defect.

  • A report saying "the local MCP server does not authenticate network callers" is not a new finding — it is already disclosed in the README and in CONTRIBUTING.md. Production wiring (network-facing endpoints with real authentication) is scheduled for the public 0.1 release.
  • A report saying "a local user can claim a higher role than they hold" is also expected: the server trusts the local caller. Only treat it as a vulnerability if a change you observed widens this boundary, e.g. a port that started binding to 0.0.0.0, or a server that began reading secrets from untrusted input.

Response expectations (conservative)

We commit to best-effort handling, not to fixed response-time numbers:

  • We will acknowledge receipt and begin triaging privately.
  • We will investigate and communicate a remediation plan or a documented "won't fix" rationale through the advisory.
  • We will coordinate any public disclosure so that a fix — or an agreed workaround — lands first, and the CHANGELOG / release notes record the issue honestly.

We are a small maintainer team, so response times vary; please do not assume silence means acceptance.

Out of scope

  • Reports against third-party services this project merely references (e.g. Claude Desktop, Cursor) — please report those to their vendors.
  • Social engineering of maintainers.
  • Issues that only affect unmodified builds run on a machine the attacker already controls with full privileges.

There aren't any published security advisories