16. Security Checklist Project

This chapter provides a practical exercise: create a security checklist for your local AI deployment. The checklist serves as both an implementation guide and an audit tool.

Build your checklist:

Step 1: Inventory assets

List every component in your AI stack: models, APIs, data stores, network connections, and dependencies. For each asset, note: sensitivity level, owner, and last review date.

Step 2: Map controls

For each asset, document which security controls apply. Use categories:

• Access control (authentication, authorization)
• Data protection (encryption, masking, retention)
• Network security (firewalls, segmentation)
• Monitoring (logging, alerting)
• Resilience (backups, redundancy)

Step 3: Define checks

Convert controls into verifiable checks. "API should require authentication" becomes "Verify unauthenticated requests return 401".

Example checklist structure:

# ai-security-checklist.yaml
meta:
  created: "2024-05-29"
  owner: [email protected]
  review_frequency: monthly

categories:
  - name: Access Control
    checks:
      - id: AC-001
        description: All AI API endpoints require authentication
        test: "curl -s http://localhost:11434/api/tags | grep -q '401'"
        frequency: weekly
        last_run: null
        status: pending
      
      - id: AC-002
        description: API keys rotate at least every 90 days
        test: "Check key creation dates in secrets store"
        frequency: monthly
        last_run: null
        status: pending
      
      - id: AC-003
        description: Principle of least privilege for service accounts
        test: "Audit service account permissions quarterly"
        frequency: quarterly
        last_run: null
        status: pending

  - name: Data Protection
    checks:
      - id: DP-001
        description: Sensitive data encrypted at rest
        test: "Verify LUKS/cryptsetup status for data volumes"
        frequency: weekly
        last_run: null
        status: pending
      
      - id: DP-002
        description: Data retention policies enforced
        test: "Check automated cleanup jobs execute successfully"
        frequency: daily
        last_run: null
        status: pending
      
      - id: DP-003
        description: GDPR/CCPA consumer request capability
        test: "Test deletion request flow in staging"
        frequency: quarterly
        last_run: null
        status: pending

  - name: Network Security
    checks:
      - id: NS-001
        description: Default deny firewall policy active
        test: "sudo iptables -L | grep -E 'Chain (INPUT|FORWARD).*policy DROP'"
        frequency: weekly
        last_run: null
        status: pending
      
      - id: NS-002
        description: No AI services exposed to public internet
        test: "nmap -Pn -p 11434,8080 YOUR_EXTERNAL_IP 2>/dev/null | grep -q '0 hosts up'"
        frequency: weekly
        last_run: null
        status: pending
      
      - id: NS-003
        description: TLS configured for all external API access
        test: "openssl s_client -connect localhost:11434 </dev/null 2>&1 | grep -q 'Protocol.*TLS'"
        frequency: monthly
        last_run: null
        status: pending

  - name: Monitoring
    checks:
      - id: MN-001
        description: Audit logs capture authentication events
        test: "grep 'authentication' /var/log/ai-audit.log | tail -1"
        frequency: daily
        last_run: null
        status: pending
      
      - id: MN-002
        description: Alerts configured for anomalous API usage
        test: "Check alert configuration against runbook"
        frequency: monthly
        last_run: null
        status: pending

  - name: Model Security
    checks:
      - id: MS-001
        description: Model integrity verified before loading
        test: "sha256sum $(ollama list --format '{{.Name}}:{{.Size}}') | compare to inventory"
        frequency: weekly
        last_run: null
        status: pending
      
      - id: MS-002
        description: No unauthorized models loaded
        test: "Compare ollama list to approved model inventory"
        frequency: weekly
        last_run: null
        status: pending
      
      - id: MS-003
        description: Prompt injection detection active
        test: "Send test injection payload, verify rejection or redaction"
        frequency: monthly
        last_run: null
        status: pending

Automate checklist execution:

#!/bin/bash
# run-security-checklist.sh

CHECKLIST="./ai-security-checklist.yaml"
REPORT="/var/log/security-checklist-$(date +%Y%m%d).log"

echo "Running AI Security Checklist - $(date)" > "$REPORT"

while IFS= read -r check; do
    id=$(echo "$check" | yq '.id')
    desc=$(echo "$check" | yq '.description')
    test_cmd=$(echo "$check" | yq '.test')
    
    echo -n "Checking $id: $desc ... " >> "$REPORT"
    if eval "$test_cmd" >/dev/null 2>&1; then
        echo "PASS" >> "$REPORT"
    else
        echo "FAIL" >> "$REPORT"
    fi
done < <(yq '.categories[].checks[]' "$CHECKLIST")

echo "Checklist complete. See $REPORT"

Conclusion

Securing local AI deployments requires deliberate architecture and ongoing maintenance. The controls in this course—authentication, network hardening, data isolation, compliance mapping, and incident response—form a foundation. Implement incrementally: start with authentication and network controls, then add monitoring, then address compliance requirements.

Security is never "done." Threat actors adapt, vulnerabilities emerge, and your deployment evolves. Build review cycles into your operations: monthly checklist execution, quarterly threat model updates, annual compliance reviews. Document everything so that when incidents occur—or when auditors arrive—you can demonstrate due diligence.

The advantage of local AI security is control. You own the infrastructure, the data, and the update cycles. Use that control to implement defense-in-depth: multiple layers that protect each other. A single control failure should not compromise your entire deployment.

Begin with the checklist project in Chapter 16. Know your current state. Identify gaps. Prioritize fixes. Repeat.