Self-Hosted Security: The 6-Layer Model

📋 At a Glance

Self-hosting means full control — but also full responsibility. Security is not a single feature but a layered model: 6 layers from physical infrastructure to monitoring. Each layer stops something different. None is sufficient on its own.

The 6 Security Layers

Each layer addresses a different attack surface. If Layer 1 fails, Layer 2 must hold. This is Defense in Depth.

Security in layers: Each level protects against different attack vectors.

Layer	Protects Against	Tools
1. Network	Unauthorized external access	Firewall (UFW/iptables), VLAN, VPN
2. SSH & Authentication	Brute force, weak passwords	SSH Key-Only, fail2ban, 2FA
3. Host Operating System	Outdated software, kernel exploits	Unattended Upgrades, CIS Benchmark
4. Containers & Services	Privilege escalation, unsecured APIs	Rootless containers, read-only FS, secrets
5. Application	Prompt injection, data leakage	Input validation, output sanitizer, rate limits
6. Monitoring & Response	Undetected intrusions	Loki, Grafana Alerts, audit logs

Layer 1: Network Segmentation

Your local AI stack should NOT be directly reachable from the internet. The most important rule: Default Deny — everything is blocked, you only open what is needed.

UFW Basic Configuration

# Block everything (Default Deny)
sudo ufw default deny incoming
sudo ufw default allow outgoing

# Allow SSH (only from local network)
sudo ufw allow from 10.40.10.0/24 to any port 22

# Reverse Proxy (HTTPS)
sudo ufw allow 443/tcp

# Enable firewall
sudo ufw enable
sudo ufw status verbose

⚠️ Do NOT expose ports publicly

Services like Ollama (11434), n8n (5678), Grafana (3000), PostgreSQL (5432) do NOT belong on the internet. If you need external access: VPN or reverse proxy with authentication (e.g., Cloudflare Tunnel or Traefik with BasicAuth).

Layer 2: SSH & Authentication

SSH is the main access point to your servers. Misconfigured, it becomes the biggest attack vector.

Hardening SSH (/etc/ssh/sshd_config)

# Disable password login
PasswordAuthentication no

# Prohibit root login (or key-only)
PermitRootLogin prohibit-password

# Allow specific users only
AllowUsers joe admin

# Idle timeout (5 minutes)
ClientAliveInterval 300
ClientAliveCountMax 0

# Restart: sudo systemctl restart sshd

Install fail2ban

# Installation
sudo apt install fail2ban

# Enable SSH jail
sudo cp /etc/fail2ban/jail.conf /etc/fail2ban/jail.local

# In jail.local:
# [sshd]
# enabled = true
# maxretry = 3
# bantime = 3600

sudo systemctl enable fail2ban
sudo systemctl start fail2ban

# Check status
sudo fail2ban-client status sshd

ℹ️ Generate SSH Keys

If you do not have SSH keys yet: ssh-keygen -t ed25519 -C "your@email.at". Copy the public key to the server: ssh-copy-id user@server. Then disable password login.

Layer 3 & 4: Host & Container Security

Host System

Measure	Command / Configuration	Why
Auto-Updates	sudo apt install unattended-upgrades	Automatically apply security patches
Non-root User	sudo adduser deploy && sudo usermod -aG docker deploy	Minimal privileges, no permanent root
Kernel Updates	sudo apt upgrade linux-generic	Close kernel exploits
Unnecessary Services	sudo systemctl disable bluetooth cups	Reduce attack surface

Container Security

Principle	Implementation	Example
Read-Only Root	read_only: true in compose	Prevents filesystem manipulation
No Root in Container	user: '1000:1000' in compose	Container runs as normal user
Secrets Management	Docker Secrets or Vault	No credentials in environment variables
Resource Limits	mem_limit: 4g, cpus: '2.0'	Container cannot overwhelm the host
Network Isolation	Dedicated Docker networks per stack	Services only see what they need

⚠️ Swagger/Docs in Production

Many frameworks (FastAPI, Express) serve API documentation at /docs or /swagger automatically. In production: disable with docs_url=None. Attackers use these endpoints to map your API structure.

Layer 5: Application Security for AI

AI applications have unique security risks. LLMs can be manipulated (prompt injection) and leak sensitive data in their responses.

Risk	Description	Countermeasure
Prompt Injection	User manipulates LLM instructions	System prompt isolation, input validation
Data Leakage	LLM outputs environment variables or secrets	Output Sanitizer (MANDATORY)
Token Theft	API keys intercepted	Vault, token rotation, rate limiting
Model Poisoning	Manipulated models from insecure sources	Only official sources (ollama.com, HuggingFace verified)
Resource Exhaustion	Excessively long prompts/contexts	Max token limits, request timeouts

⚠️ Output Sanitizer is Mandatory

LLMs can be tricked into outputting environment variables, API keys, or system information. EVERY response must pass through a sanitizer that detects and removes patterns like API keys, IP addresses, and file paths. This is not a nice-to-have — it is mandatory for production.

Layer 6: Monitoring & Incident Response

Security without monitoring is blind. You need to know what happens on your systems — in real time.

What to Monitor	Tool	Alert Condition
SSH Login Attempts	fail2ban + Grafana	More than 5 failed logins/hour
Container Health	Docker Health Checks + Uptime Kuma	Container unhealthy or stopped
Disk Usage	Prometheus Node Exporter	Disk > 85% full
API Errors	Loki Log Queries	HTTP 5xx rate > 1% of requests
Unexpected Processes	Audit Logs (auditd)	New process from unknown user

💡 Security Dashboard in Grafana

Create a dedicated security dashboard with fail2ban bans, SSH logins, API error rates, and disk usage at a glance. You can see in 5 seconds if something is off. Our Grafana Monitoring Guide covers setting up security panels.

Das Wichtigste

✓Security is a layered model. 6 layers, each stops something different. None is sufficient alone.
✓Default Deny at the firewall. Only open what is needed. AI ports do NOT belong on the internet.
✓SSH Key-Only + fail2ban. No password login, automatic ban after 3 failed attempts.
✓Output Sanitizer is mandatory for AI applications. LLMs will leak secrets and system info otherwise.
✓Monitoring with alerting. Without monitoring, you will not know if someone is already inside.

Sources

CIS Benchmarks — Industry standard for OS hardening
Docker Security Documentation — Container security best practices
OWASP Top 10 — Most common web application security risks
fail2ban — Brute force protection for SSH and other services

Security