Skip to content

Security: synapseorch-ai/synapse-ai

SECURITY.md

Security Policy

Reporting a vulnerability

Please do not open a public issue for security problems.

Report privately through GitHub Security Advisories, which creates a private thread visible only to the maintainers.

If you cannot use GitHub advisories, email info@synapseorch.com with SECURITY in the subject.

Please include:

  • what the issue is and why it is a problem
  • the version or commit you tested (synapse --version)
  • how the instance was deployed — local, Docker, or scale mode
  • steps to reproduce, and a proof of concept if you have one

What to expect

First response within 5 working days
Assessment and severity within 10 working days
Fix and coordinated disclosure agreed with you, normally within 90 days

We will credit you in the advisory unless you would rather stay anonymous. We do not run a paid bounty programme.

Supported versions

Security fixes land on the latest minor release. Older versions are not backported — please upgrade.

Deployment guidance

Synapse runs AI agents that execute tools on your machine. Most reported issues come from exposing an instance that was configured for local use, so a few defaults are worth knowing.

Do not publish the backend port on a public interface. The backend (default 8765) hosts the internal API and the agent chat endpoints. Reach it through the frontend on port 3000 instead. Docker Compose binds the backend to 127.0.0.1 by default; SYNAPSE_BACKEND_BIND overrides this, and changing it exposes those routes to your network.

Set SYNAPSE_INTERNAL_TOKEN. It gates the internal /api/* and /chat routes. Docker images generate one automatically on first boot. Without it, those routes are restricted to loopback callers, so a network-exposed instance without a token will reject remote requests rather than serve them.

Enable login (Settings → General) on anything reachable beyond your own machine.

Treat these as remote code execution by design, and only enable them where you trust every user who can reach the instance:

  • stdio MCP servers — registering one runs a local command. Disable with allow_stdio_mcp, or restrict with mcp_command_allowlist. Force-disabled in scale mode.
  • Custom Python tools and the transform step — run customer-authored Python. The Docker runtime sandboxes them. A transform runs with no network (since 2.0): it reaches outside through a tool step. A transform's code from an imported bundle is listed in the import's preview and warnings.
  • SYNAPSE_TRANSFORM_RUNTIME=host — set only in the server's environment, never through settings (POST /api/settings refuses it). It runs transforms outside the sandbox, as the server, able to read everything the server can, including stored keys, with the network; their environment is cut to a short allowlist (no tokens, keys, secrets or URLs). Only for a single-user machine you trust every workflow on.
  • The bash tool and execute_python — run commands on the host or in a container.

Agents are subject to prompt injection. Any content an agent reads — a web page, a document, a tool result — can attempt to steer it. Give agents the narrowest tool set that does the job, and prefer human-approval steps for anything irreversible.

API keys (sk-syn-…) carry full access to the v1 and v2 APIs. Treat them as passwords, and rotate them from Settings → API Keys if one leaks.

Past advisories

Published advisories are listed under Security → Advisories.

Learn more about advisories related to synapseorch-ai/synapse-ai in the GitHub Advisory Database