Open a private security advisory on the repository, or contact the copyright holder named in
LICENSE. Do not open a public issue for anything that could let another tenant read agent data or
forge telemetry. Expect a first response within a week.
AgentMiniOffice listens on the public internet behind a domain, so assume it will be found and probed.
| Asset | Risk | Mitigation |
|---|---|---|
| Agent metadata and task titles | Readable by anyone who leaks a share link | Share links are read-only, scoped, expiring, revocable |
| Ability to publish fake telemetry | Office shows fiction | Agent tokens are per agent, hashed at rest, revocable, scope-limited to ingest and inbox |
| Prompt and file content | Would be a trove on a public host | Not collected. Server rejects content-looking payloads |
| Control of agents | Attacker pushes tasks | Downlink is pull only, with lease and ack, no shell execution ever |
| Admin UI | Full visibility of all agents | Separate user sessions with httpOnly cookies, roles, and no reuse of agent tokens |
- AgentMiniOffice does not execute commands on agents. It stores tasks. The agent decides what to do.
- AgentMiniOffice does not store model output. If you need that, use a dedicated tracing backend.
- AgentMiniOffice is not hardened for multi-tenant hostile isolation in v1. One organization per deployment is the supported configuration until v3.
Keep office enrollment separate from private machine and human credentials:
-
amo_join_*is a reusable office enrollment key. Only that office's Owner can reveal or rotate it, even while Closed. It cannot report for a worker. Closing blocks new enrollment; revoke an installation to stop its existing access. -
The collector derives each issued office key from a per-office nonce and a random server-only secret stored next to SQLite as
<db-file>.join-secret(0600 on POSIX). Keep the database directory private. Back up the sidecar separately with restricted access, and restore it with the corresponding database. A DB-only copy cannot reveal office keys. Losing the sidecar prevents reveal, but does not invalidate existing verifiers: restore the file or explicitly rotate the affected keys. Do not ship either file in an export. -
amo_producer_*andamo_writer_*credentials remain private, hash-only on the collector. There is no reveal endpoint for them. -
amo_agent_*ingest tokens. Stored as SHA-256 hashes plus an eight character prefix for display. Shown exactly once at creation. Limit: ingest and inbox for one agent id. -
Session cookies. HttpOnly, SameSite=Lax, server-side session rows, revocable.
-
Share links. Signed, read-only, expiring, revocable.
The token that an agent holds must never be able to open the panel, and a logged in user must never be able to forge events for an agent. Both directions are enforced by separate middleware.
- TLS is terminated by a reverse proxy (Caddy is the recommended default).
- Bind the Node process to localhost only. Never expose the port directly.
- Agent tokens are 32 random bytes, stored as SHA-256 hashes. A database leak does not hand out working tokens, but it does leak agent names, task titles, and timings. Treat the database as sensitive.
- Enable rate limiting per token before the collector is reachable from the internet. A runaway agent in a loop is the most likely denial of service against your own panel.
Events are pruned after 14 days. Daily aggregates survive. Back up via sqlite3 .backup or Litestream,
never by copying the .db file while WAL is active.