Security Model

Privilege separation

• fyrwall-server: unprivileged. Never runs firewall commands.
• fyrwall-agent: unprivileged. Collects state and events.
• Privileged helper: tiny root daemon, Unix socket 0660, strictly

allowlisted typed operations, revalidates every request, SO_PEERCRED

peer checks. No generic exec API exists anywhere.

Authentication

Argon2id password hashing (constant-time verify), server-side sessions

with secure cookies, CSRF tokens on all mutations, per-IP login rate

limiting, session revocation on password change.

Transport

TLS 1.2+ when enabled; security headers always (CSP frame-ancestors

none, nosniff, no-referrer). Agent traffic is mutually authenticated.

Domain binding rejects foreign Host headers.

Logs are protected

Server logs: 0640, service-group readable only. Agent logs: AES-256-GCM

encrypted at rest with a host-local 0600 key. Only the web admin can

read logs through the authenticated API. Retention bounds every log by

size (50 MiB x 5 backups x 30 days default) so disk can never fill.

Secrets

Never logged, never in diagnostics bundles, never on argv. Passwords

and tokens travel via environment or secret files (env:, file: refs).

What extensions can never do

No execution, no direct firewall mutation, no user management, no

credential or key access. Extensions are declarative and

capability-scoped, denied by default, granted by an admin at install.