Skip to content

security.md

Latest commit

 

History

History
161 lines (120 loc) · 7.15 KB

security.md

File metadata and controls

161 lines (120 loc) · 7.15 KB
Raw

Security — RemoteVibeServer

This document describes the security model, threat mitigations, and best practices implemented in RemoteVibeServer.

Secrets Handling

Principle: No Secrets in the Repository

All credentials and API tokens are provided exclusively at deploy-time via the cloud-init user-data. The repository contains only:

  • Example files with clearly marked <PLACEHOLDER> values
  • Scripts that read secrets from the runtime environment file

Runtime Secret Storage

Artifact Path Permissions Owner
Environment / secrets file /etc/dev-server/env 0600 root
Coder admin API token /etc/dev-server/coder-admin-token 0644 root
Agent environment profile /etc/profile.d/agent-env.sh 0644 root
Deployment status /etc/dev-server/status 0644 root
Provisioning log /var/log/dev-server-provision.log 0644 root

The environment file is:

  • Written by cloud-init before any script executes
  • Readable only by root (mode 0600)
  • Never exposed over the network
  • Never logged or printed by any script

Secret Lifecycle

Cloud Provider Console / API
        │
        ▼
  cloud-init user-data  ──►  /etc/dev-server/env  (0600 root:root)
        │                            │
        │                            ├──► bootstrap.sh (source)
        │                            ├──► setup.sh     (source)
        │                            └──► systemd EnvironmentFile
        │
        ╳  Secrets never written to stdout, logs, or /tmp

API Token Management

Cloudflare API Token

  • Must be a scoped API token (not a Global API Key)
  • Required permissions: Zone → DNS → Edit for the target zone only
  • The token is used once during provisioning to create/update the DNS A record
  • It remains in the env file for potential re-runs but is never exposed

Best practice: Create a dedicated API token in the Cloudflare dashboard → My Profile → API Tokens → Create Token → Edit zone DNS.

AI Agent API Keys

Key Used By Scope
GITHUB_TOKEN Copilot CLI, Codex read:user, copilot
ANTHROPIC_API_KEY Claude Code CLI API access
GOOGLE_API_KEY Gemini CLI Generative AI API
OPENAI_API_KEY Codex CLI, OpenCode OpenAI API access
CODEX_OPENAI_AUTH_CODE Codex CLI OpenAI OAuth (ChatGPT plan)
  • Keys are loaded into the shell environment via /etc/profile.d/*.sh scripts
  • Each profile script reads from the env file — it does not hardcode values
  • If an agent is disabled (ENABLE_AGENT_*=false), its key is ignored entirely

HTTPS & Certificates

Automatic TLS via Caddy

Caddy handles the entire TLS lifecycle:

  1. Certificate issuance — via Let's Encrypt ACME HTTP-01 challenge
  2. Certificate renewal — automatic, ~30 days before expiry
  3. Certificate storage — Caddy's data directory (/var/lib/caddy/)
  4. Protocol — TLS 1.2+ (TLS 1.3 preferred)

Security Headers

The Caddyfile injects the following headers on every response:

Header Value
Strict-Transport-Security max-age=63072000; includeSubDomains; preload
X-Content-Type-Options nosniff
X-Frame-Options SAMEORIGIN
Referrer-Policy strict-origin-when-cross-origin
Server (removed)

Why Not Certbot + Nginx?

Aspect Caddy (chosen) Certbot + Nginx
Certificate mgmt Fully automatic Cron job + hooks
Renewal failures Self-healing Silent failure risk
Config lines ~20 ~60+
WebSocket config Automatic Manual proxy_pass config

Network Security

Firewall (UFW)

Default incoming:  DENY
Default outgoing:  ALLOW
Port 22/tcp:       ALLOW  (SSH)
Port 80/tcp:       ALLOW  (HTTP → HTTPS redirect)
Port 443/tcp:      ALLOW  (HTTPS)

All other inbound traffic is silently dropped. Coder listens on 127.0.0.1:3000 and is never directly accessible from the internet.

Brute-Force Protection (fail2ban)

fail2ban is enabled with default settings to protect SSH:

  • Monitors /var/log/auth.log
  • Bans IPs after 5 failed login attempts
  • Ban duration: 10 minutes (progressive)

SSH Hardening (Recommended)

The provisioning scripts do not modify SSH configuration to avoid locking out the operator. However, we strongly recommend:

# /etc/ssh/sshd_config additions
PermitRootLogin prohibit-password
PasswordAuthentication no
PubkeyAuthentication yes
MaxAuthTries 3

Docker Security

  • Docker is installed from the official Docker repository (not distro packages)
  • Coder workspaces run as unprivileged containers by default
  • The workspace image uses codercom/enterprise-base which runs as a non-root user

Audit & Observability

Log Contents
/var/log/dev-server-provision.log Full provisioning output
/var/log/caddy/access.log HTTPS access logs (rotated)
journalctl -u coder Coder server logs
journalctl -u caddy Caddy proxy logs
/var/log/auth.log SSH authentication (fail2ban)

Threat Model Summary

Threat Mitigation
Secrets in source control All secrets injected via cloud-init at deploy-time
Unencrypted traffic Caddy enforces HTTPS with HSTS
SSH brute-force fail2ban + key-only auth (recommended)
Unauthorized port access UFW denies all except 22, 80, 443
Certificate expiry Caddy auto-renews certificates
Env file exposure Mode 0600, root-only access
Compromised workspace container Unprivileged containers, no host mounts by default