Local AI Security Hardening Guide

# Local AI Security Hardening Guide

You are a security engineer specializing in local AI deployments. Produce a hardening guide for the described setup.

## Deployment Details

- Stack: [STACK, e.g., "Ollama + Open WebUI on Ubuntu 22.04"]
- Users: [WHO HAS ACCESS, e.g., "10 internal employees via LAN"]
- Data sensitivity: [SENSITIVITY LEVEL]
- Exposure: [LOCAL NETWORK ONLY / VPN / Internet-facing]
- Known threat model: [LIST CONCERNS, e.g., "insider threat", "stolen laptop", "prompt injection attacks"]

## Hardening Checklist

### Network Isolation
Steps to restrict the inference server to localhost or specific subnet, including firewall rules.

### Authentication & Authorization
Recommended auth mechanism for [STACK] (API key, OAuth proxy, or mTLS). Provide the config snippet.

### Prompt Injection Defenses
3 techniques to detect and block prompt injection in the system prompt layer and at the application level.

### Model & Weight Protection
How to restrict file system access to model weights so only the inference process can read them.

### Audit Logging
What to log (prompt hash, user, timestamp, token count — not full prompt content for sensitive use cases), where to store logs, and how long to retain them.

### Incident Response
A 5-step response playbook if the local AI system is compromised or leaks data.