Skip to content

Security: pandudpn/ompwire

SECURITY.md

Security

Reporting

Do not open a public issue for a security problem.

Use GitHub's private report: Security tab, then Report a vulnerability on github.com/pandudpn/ompwire. If that is not possible, email pandu@pandudpn.id with ompwire security in the subject.

Expect an acknowledgement within a few days. This is a single-maintainer project with no on-call rotation, so there is no promised fix window. You will get a straight answer about whether the report is in scope, and credit in the advisory unless you ask otherwise.

Supported versions

The newest release tag only. install.sh resolves that tag by default and older tags are not patched. main is unreleased code.

Threat model

ompwire exposes a personal agent on your own machine to the chat platforms you connect as gateways, and to your own terminal through ompwire chat. It is single user on purpose, and the boundaries follow from that. Every gateway gets the same rules below; Telegram and WhatsApp are the ones implemented today.

Who can talk to the agent

  • Every update is dropped unless the sender's user id is in the gateway's allowed_users (a Telegram user id, a WhatsApp phone number). Strangers get no reply, not a refusal; the reason goes to the journal only.
  • Groups (Telegram and WhatsApp): only listed groups are read, only allowed members in them are answered, and the group's mode limits it further (mention by default, read never answers). Being in a group with the bot gives a stranger nothing.
  • Telegram channels are the one exception to the allow list: a post has no sender, so a listed channel in mode all runs every post, and anyone who can post there drives the agent. List only channels whose admins you trust as much as an allowed user.
  • Groups and channels are quiet by default: other members see the answer, not the tool trace (which can hold paths, command output and file contents) unless /verbose turns it on.
  • No inbound network. The Telegram gateway uses long polling and the WhatsApp gateway an outgoing websocket: no port, no webhook secret, nothing listening. A future gateway that needs a webhook is a change to this model and has to be documented here.
  • ompwire chat is reachable only by whoever already has a shell as your user.

What the agent may do without a person

  • approval_mode = write: shell commands that can change the machine need a button press.
  • Read-only commands pass without asking. The command is tokenised, a pipeline passes only when every part is read-only, and a redirect only into /dev/null. Flags that make a reader write (git --output, sort -o, rg --pre, ...) are refused.
  • An unanswered request is denied after [approval] timeout.
  • Scheduled runs answer cancelled to every permission request unless the schedule opts into yolo.

What leaves the machine

  • Outgoing files must resolve inside the agent's working directory, exist, fit files.max_mb, and at most five go out per reply. A model that read ~/.ssh cannot post it by naming it.
  • Nothing is sent anywhere except the connected chat platforms and the model provider OMP talks to. No telemetry.

Secrets

  • Each gateway's token lives in its own ~/.config/ompwire/gateways/<name>/.env (mode 0600), loaded by that gateway's unit through EnvironmentFile. A gateway's process never holds another gateway's token, and tokens are never written into a unit, the config or the repo.
  • The agent process starts without the harness's secrets: child_env() removes every bot token (TELEGRAM_BOT_TOKEN, every OMPWIRE_GATEWAY_*_TOKEN) and memory API keys before spawning omp, schedules included.
  • A WhatsApp gateway has no token. Its credential is the linked-device session in ~/.local/state/ompwire/gateways/<name>/whatsapp.db, created 0600 before whatsmeow opens it; nothing logs its contents, and the agent process never gets its path.
  • Logs mask token-shaped strings before they are written; ompwire scrub-logs cleans older files.
  • ompwire holds no provider keys. OMP authenticates itself.

What ompwire touches in ~/.omp

  • It reads OMP's session transcripts to compute /usage totals and compaction notices.
  • /compaction edits only the compaction: block of ~/.omp/agent/config.yml, atomically, after a one-time backup.
  • uninstall never touches ~/.omp.

In scope

  • A token, chat id or other secret written to a log, unit file, error message or crash trace.
  • An allow-list bypass: any path where an update from a user not in allowed_users reaches the agent, including a group or channel that is not listed, a member who is not allowed, or a group message in mention mode that did not address the bot.
  • A message the bot sends that mentions (tags) someone the agent did not name.
  • An approval bypass: a command the read-only check lets through that can write, delete or run code.
  • A <<file:>> path that escapes the working directory or the limits.
  • A harness secret that reaches the agent process.
  • Command injection or path traversal through configured values: cwd, home_channel, gateway names, MCP entries in omp's mcp.json written by setup, schedule names.
  • Markup the agent did not intend that renders as HTML or a link usable against the reader.
  • A dependency or install-script flaw that gives remote code execution.

Out of scope

  • What the agent does when you ask it to. It runs with your shell, files and credentials by design; approvals and yolo are the control, and that control is yours.
  • Anyone you put on the allow list. Adding a user id hands over the agent.
  • Denial of service against your own machine: a runaway turn, a huge context, a model bill.
  • Local access to a machine where the config and .env are readable.

If a token leaks

Revoke first, then repair the local copy. For Telegram:

# 1. revoke in @BotFather, then
ompwire gateway start telegram           # answer yes to "enter a new one?"; checked with getMe
ompwire scrub-logs                       # clean any log file the old token reached

If the token reached a git history you push, treat the bot as compromised, not only the token: whoever used it could read your conversations and act as you until it was revoked.

For WhatsApp the credential is the session file ~/.local/state/ompwire/gateways/<name>/whatsapp.db: whoever copies it acts as the linked device. Remove the device on the phone (Settings > Linked devices), then ompwire gateway start <name> --relink.

There aren't any published security advisories