Security model

C3 Mobile's core promise: your hub and network never see control content. Prompts, terminal output, and commands are protected end to end between your phone and each bridge. Push is deferred from the current beta; its planned payload contains only an opaque identifier and generic attention text.

NOTE — > C3 Mobile is beta software and its protocol has not yet completed independent cryptographic review. The design is documented and testable, but treat "end-to-end encrypted" as an engineering claim, not an audited guarantee.

Message protection

Every command and event travels as a sealed envelope:

XChaCha20-Poly1305 authenticated encryption, with keys derived per phone↔bridge pair (X25519 key agreement + HKDF-SHA256).
Context binding — the authenticated data ties each packet to its exact message subject, protocol version, hub/host/phone identities, channel epoch, direction, kind, and sequence number. A valid ciphertext replayed on a different subject, session, or position is rejected.
Fresh channels — each connection derives ephemeral channel keys, re-keyed on reconnect and every 15 minutes; command and event directions use independent keys, and sequence counters are monotonic with replay windows.
Padding — ciphertext is padded into coarse size buckets so lengths leak little about content.
Expiry — inputs expire in ~5 seconds, mutations in ~10, reads in ~30. Delayed messages die instead of executing late.

Establishing trust

Transport pinning — the phone validates your hub's CA public key (SPKI SHA-256) inside a native module on iOS and Android, and refuses any other certificate, host, or port. There is no "accept anyway" path. Bridges pin the same CA over TLS 1.3.
Enrollment — the hub-side QR carries a short-lived secret; the phone generates its keys locally, proves possession, and receives its own signed NATS credential. Private keys never leave the device.
Pairing — phone and bridge run a mutual key-confirmation exchange over a one-time QR (≤10 minutes validity), deriving a long-term pair root the hub never learns. Each pair has independent keys and counters.

Secrets at rest

Phone: hardware-backed secure storage; Android sets FLAG_SECURE (no screenshots/recents preview), iOS covers the app switcher; opening sessions and terminals sits behind your device biometric/passcode.
Bridge: OS credential store by default (C3BRIDGE_SECRET_BACKEND=keyring); explicit encrypted-file backend for headless hosts. State on disk is encrypted; logs carry event codes only — never prompts, paths, subjects, or tokens.
Hub: administration keys live in the hub machine's keystore and in the passphrase-sealed backup you create with hub control backup. The deployable bundle contains no secrets.

Revocation

Two independent layers, both effective while the target is offline:

Unpair — from the app (Settings → paired bridges) or c3bridge revoke <device-id> on a host: removes that one phone↔bridge relationship.
Hub revocation — c3sync hub control revoke <USER_PUB>: invalidates the device's hub credential via a signed, monotonic revocation manifest; bridges consume it on their next connection. Use this for a lost phone — see the ceremonies.

What this does NOT protect against

Honest non-goals, from the threat model:

A compromised endpoint. A rooted/compromised phone, bridge host, or Herdr install can read whatever that endpoint can read. E2EE protects the path, not the ends.
Recalling data. Revocation stops future access; it cannot erase what a device already received.
Traffic analysis. The hub operator (you, usually) sees subjects, timing, and coarse sizes. There is no cover traffic.
Denial of service. Whoever controls the network or the NATS server can always drop or delay messages — expiry turns that into failed commands, not wrong ones.
Permission-prompt semantics. Allow/Reject in the cockpit sends keystrokes (Enter/Esc) to the agent's terminal; C3 does not parse or verify the prompt the agent is showing. Glance at the terminal when the stakes are high.

Related: architecture · hub commands & ceremonies · C3Sync's separate security model.