Skip to content

Security: igorrius/ganglion

SECURITY.md

Security policy

ganglion runs code a model wrote, runs commands on its host and reads text from strangers. Its design rule is that every link is hostile and every input is untrusted, and that what stops an abuse is code, not a prompt (spec, principle 9). A report that shows where the code does not hold that line is the most useful one the project can get.

Reporting a vulnerability

Report it privately, through GitHub's private vulnerability reporting ("Report a vulnerability" on the repository's Security tab). Do not open a public issue for it.

Say what you did, what happened and what should not have: a configuration, a message or a module that reproduces it is worth more than a description. Leave out real tokens and the address of any live node.

The project is 0.x and has one maintainer, so there is no response time to promise and no bounty. You will get an answer, a fix on master when there is one, and credit in the advisory if you want it. There are no supported old versions: a fix lands on master only.

In scope

  • The module sandbox. A WASM module that reaches anything its manifest did not ask for and the owner did not approve: a host, a file, a secret, another module's state, more memory or time than its limit.
  • The command sandbox. A host command or a skill's script that escapes its restrictions: reads the node's database, its key or its secrets, gains privileges, or reaches what its class does not allow.
  • Secrets. A secret that appears in a log, the journal, a metric, an error a model reads, a message, or a child process's arguments; a weakness in the encrypted store.
  • Prompt injection that crosses a boundary in code. Outside text (a web page, a file, a message in a group, a module's output) that makes the node run a state-changing action without the owner's confirmation, send private data out, or change its own configuration, memory or skills.
  • Access. A person who is not on the access list getting an answer, a member of a group getting an admin's rights, a forged webhook call that is accepted.
  • The node's listeners and the diagnostics server (ganglion mcp).
  • The release path of modules: a module accepted with a wrong checksum or with rights above the node's ceiling.

Out of scope

  • A model saying something wrong or rude with no boundary crossed.
  • What the owner approved with a button, or what someone can do who already has the node's files, its shell or its admin role.
  • Parts the spec marks as not built yet.
  • Denial of service that needs an account on the access list.
  • Vulnerabilities in dependencies with no path to exploit them through ganglion: report those upstream.

Running a node safely

Run it as an unprivileged user, give it its own bot token and its own provider keys, and keep its data directory private: it holds the conversations, the memory and the key of the secret store.

There aren't any published security advisories