Hardening an AI Development Environment

February 7, 2026 security, infrastructure, wsl2, ai-agents, devops

Changes: v1.2 → v1.3

Architecture layer added: Security. Six components deployed, four automated daily checks registered.

Components Added

Component	Purpose	Install	Status
fail2ban	Intrusion prevention — bans IPs on auth failure patterns	`apt install fail2ban`	`systemctl is-active` ✅
ufw	Firewall — default deny incoming, allow outgoing	`apt install ufw && ufw enable`	`ufw status` ✅
AppArmor	Mandatory access control — confines process resource access	`apt install apparmor apparmor-utils`	`aa-status` ✅
unattended-upgrades	Automated security patches	`apt install unattended-upgrades`	`systemctl is-active` ✅
rkhunter	Rootkit/binary integrity scanning	`apt install rkhunter`	`rkhunter --version` ✅
Security check registry	4 daily automated checks (WSL config, file perms, package updates, Docker bloat)	Custom — `~/.claude/security-improvements/registry.json`	Active ✅

Daily Automated Checks

WSL configuration validation — systemd settings, user permissions
File permission auditing — sensitive path monitoring
Package update monitoring — pending security update detection
Docker bloat detection — container disk usage tracking

Audit Pipeline: `audit_map.py`

Bidirectional verification — blocks production deploy on any failure.

Phase 1 — Claims → Reality: Every architecture component verified by shell command. Example: systemctl is-active fail2ban. Fail = flagged fabricated.

Phase 2 — Reality → Claims: Discovers running services, flags undocumented ones.

Phase 3 — Changelog consistency: Version entries validated. Referenced blog posts checked for existence and non-draft status.

Phase 4 — Stack versions: Declared versions compared against installed versions.

Current Results

✅ Passed:       31
❌ Failed:       0
🚫 Fabricated:   0
⚠️  Undocumented: 0

Rationale

AI agents in this stack have filesystem, shell, and network access. Agent-level governance (task registry, permission allowlists, pre-execution hooks) covers the application layer. OS-level hardening covers everything below it. Defense in depth.

Architecture version: v1.2 → v1.3. All claims verified by scripts/audit_map.py.

Configuration details reflect a production environment at time of writing. Implementation specifics vary based on tooling versions, platform updates, and organizational requirements. Validate approaches against current documentation before deployment.

← Back to Journal