Ujex security model

Three-layer security

1. Firestore Security Rules — ownership and tenant isolation
2. Cloud Function scope checks — capability-token-based authorization on every callable
3. Append-only hash-chained audit ledger — post-hoc verifiability

Identity

Every agent has a Firebase UID-equivalent. Owners (humans) own agents. Device keys are stored as sha256(deviceKey) — never plaintext. Scopes (postbox, recall, mesh, scheduler, etc.) are granted per device key.

Secrets

KMS-envelope encryption for: device keys (hashed), agent secrets, tool auth tokens, DDNS tokens, gateway certs. Cloud KMS keys never leave KMS.

Mail security

• Inbound MIME parsed, HMAC-signed at the bridge, verified at the Cloud Function
• Prompt-injection scored (heuristic regex + Gemini fallback)
• Outbound DKIM/SPF/DMARC signed (via OpenDKIM and/or Brevo's selectors)

Audit ledger

Every privileged action: audit/{seq} with sha256(prev || event). Hourly verification by Cloud Scheduler. Details.

Threat model

In scope:

• Stolen device key → revoke via callable, audit log records the rotation
• Compromised owner password → Firebase MFA + revoke-all-tokens via Admin SDK
• Prompt injection in inbound email → PI score surfaced; agent routes high-PI to human approval
• Tampering with audit log → hash chain breaks at modified row, verifier alerts

Not fully in scope:

• Compromise of the underlying Firebase project (game over)
• Sophisticated prompt injection that scores low (we score; we don't claim to catch everything)
• Compromise of Cloud KMS (which would mean Google's KMS is compromised)

Reporting

Email security@ujex.dev. We respond within 48h.