infra-security-audit

Infra Security Audit

Overview

This skill audits and hardens the host and network layer of a self-hosted Linux server (firewall, SSH, intrusion blocking, exposed ports, public-exposure posture). It is distinct from agent/application-layer security (skills, permissions, sandboxing). It is built for non-developer operators running a single VPS or home server, and it never makes a risky change on its own.

The single most important rule: audit reads, applying changes is gated. Diagnosis is automatic; anything that can sever a connection, lock the current session, or stop a service requires explicit human approval, a backup, and a self-lockout safeguard.

Workflow Decision Tree

Operate in three modes, always in this order. Never skip from audit straight to applying risky changes.

1. audit (default, read-only) — Diagnose and report. Safe to run anytime.
2. plan — For each finding, classify the fix as GREEN / YELLOW / RED and produce a concrete diff/command preview. No changes yet.
3. apply — Execute GREEN fixes automatically; for YELLOW/RED, show the diff, take a backup, get explicit approval, apply, then verify (and roll back on failure).

When the user only says "점검" or "audit", stop after mode 1 and report. Only proceed to plan/apply when the user asks to fix or harden.

Mode 1 — audit (read-only)

Run the bundled audit script. It is read-only and safe.

bash scripts/audit.sh            # local host
bash scripts/audit.sh user@host  # remote over SSH (key-based)

The script checks the items in references/checklist.md, prints a hardening score (0–100) and a two-tier report ([즉시] must-fix / [권장] recommended), and labels each finding [자동] (safe to auto-fix) or [승인] (needs approval). If a check cannot run (tool missing, permission denied), it prints 미확인 rather than guessing.

To interpret raw command output (ufw status, sshd -T, fail2ban, ss, Lynis/ssh-audit if present), load references/interpret.md.

Reuse existing tools, do not reinvent. If lynis or ssh-audit is installed, the script uses them and filters to this skill's scope; otherwise it falls back to direct checks. See references/checklist.md.

Mode 2 — plan

For each finding from the audit, classify the remediation by risk and produce a preview. Load references/safe-boundary.md for the full GREEN/YELLOW/RED table. Summary:

• GREEN — non-destructive, auto-apply OK: does not cut connectivity, lock the session, or stop a service. Examples: install fail2ban + enable sshd jail (with operator IP/Tailscale range added to ignoreip), enable unattended-upgrades, add operator IP to an allowlist, add a validated Caddy rate_limit and reload.
• YELLOW — reversible, needs approval: sshd hardening (PermitRootLogin, PasswordAuthentication, weak algorithms), sysctl network params.
• RED — hard to reverse, hard gate + mandatory backup: ufw enable, SSH port change, binding a port from 0.0.0.0 to loopback/Tailscale, switching to CrowdSec, enabling auto-reboot.

Produce, for every non-GREEN fix, a diff or exact command preview ("this file changes like this") before doing anything.

Mode 3 — apply

• Execute GREEN fixes, then re-audit to confirm the score moved.
• For YELLOW/RED: show diff → take a timestamped backup (e.g. sshd_config.bak.<ts>) → get explicit approval → apply → verify in a new session → roll back on failure.
• Self-lockout prevention is mandatory. Before any firewall or SSH change, load references/self-lockout.md and follow it exactly. Core safeguards: allow SSH (and the Tailscale range) in ufw before ufw enable; put sshd changes in a low-numbered drop-in and verify a fresh login keeps working before committing; check /etc/ssh/sshd_config.d/*.conf ordering for the first-match trap.

Safety Rules (always)

1. Stop on empty/ambiguous results. If an audit check returns nothing or errors, never make a guessed change — report and escalate to the human.
2. Back up before any file change. Keep a one-command rollback path.
3. Log every command and its result so the run is auditable.
4. Idempotent. Re-running must not double-apply.
5. Never auto-run a RED change, even if the user says "fix everything" — RED always shows diff + backup + explicit confirm.
6. This skill is itself third-party code running with the agent's privileges; the human approval gate is the only real boundary for risky changes. Treat it as load-bearing.

Resources

• references/checklist.md — the 14 scoped audit items (P1–P3) and which tool covers each.
• references/safe-boundary.md — GREEN/YELLOW/RED classification and the rationale (cut/lock/stop test).
• references/commands.md — copy-paste remediation commands, dogfood-verified (ufw, SSH 00-hardening drop-in, fail2ban, Caddy rate_limit, Cloudflare, Tailscale).
• references/self-lockout.md — the three self-lockout traps and how to avoid each.
• references/interpret.md — how to read ufw/sshd -T/fail2ban/ss/Lynis/ssh-audit output.
• scripts/audit.sh — read-only audit engine (local or user@host), prints score + report.

Tone (when reporting to a non-developer)

Plain Korean, calm, non-fear-mongering. Frame attacks as "the internet's default weather", not a targeted threat. Do not assert attacker nationality (bots spoof and relay; origin is meaningless). Do not name specific vendors as "bad" or cite CVE numbers as fact. No emoji in output — use text labels [즉시] / [권장] / [자동] / [승인].