Security Audit — Three-Layer Linux Security

Covers three security layers:

1. Host/Hardware — fwupd HSI, TPM, Secure Boot, kernel taint, power management
2. System/Software — ports, services, permissions, firewall, users
3. Local AI Services — Ollama install/removal, config sync, network exposure

Layer 1: Host Security (Hardware + Firmware)

See references/host-security-audit.md for full guide.

Quick Reference

set -euo pipefail
# HSI level
fwupdmgr security 2>&1 | head -30

# Kernel taint
cat /proc/sys/kernel/tainted

# Secure Boot
mokutil --sb-state

# TPM
ls /sys/class/tpm/tpm*

# Lockdown
cat /sys/kernel/security/lockdown

HSI Levels

Level	Meaning	Common Items
HSI-1	UEFI Secure Boot, TPM 2.0, ME locked	UEFI PK, Boot Services Vars, MEI
HSI-2	Intel BootGuard, IOMMU	BootGuard Fuse, IOMMU, Platform Debug
HSI-3	Suspend, DMA Protection, CET	Suspend To Idle, Pre-boot DMA
HSI-4	Maximum (server-class)	Encrypted RAM (TME/MKTME)

Suspend To Idle Fix (HSI-3)

Four configs needed:

1. Kernel: mem_sleep_default=s2idle in cmdline
2. systemd: /etc/systemd/sleep.conf.d/s2idle.conf with SuspendState=mem
3. logind: /etc/systemd/logind.conf.d/s2idle.conf with HandleLidSwitch=suspend
4. GNOME: gsettings set org.gnome.settings-daemon.plugins.power lid-close-suspend-with-external-monitor true
5. Refresh: kill -USR1 $(pgrep gsd-power)

Kernel Taint Triage

Bit	Name	Common Cause
0	PROPRIETARY_MODULE	NVIDIA, ZFS, VirtualBox
12	STALE_HW	ACPI BIOS errors, firmware-DMI mismatch

Layer 2: System Security (Software)

See references/system-security-audit.md for full guide. See references/gcp-ubuntu-hardening.md for Google Cloud Ubuntu VM hardening: OS Login, restrictive GCP firewall + UFW, Shielded VM, encryption model decisions, Ops Agent, snapshots, and performance baseline before tuning.

Quick Reference

set -euo pipefail
# Firewall + ports
sudo ufw status
ss -tlnp | grep -E "0\.0\.0\.0:|:::"

# Permissions
ls -la ~/.hermes/state.db ~/.hermes/config.yaml ~/.ssh/
ls -la ~/.hermes/logs/ ~/.hermes/state-snapshots/

# Services
systemctl list-units --type=service --state=running | head -30

# Users
sudo cat /etc/passwd | grep -E "/home|/bin/bash" | cut -d: -f1,3,7

# Updates
apt list --upgradable 2>/dev/null | grep -c "/"

Quick-Fix Cheatsheet

set -euo pipefail
# Dienst deaktivieren
sudo systemctl stop <dienst> && sudo systemctl disable <dienst>

# Häufige Dienste mit Sicherheitsrisiko
sudo systemctl disable --now gnome-remote-desktop.service  # RDP auf *:3389
sudo systemctl disable --now zramswap.service              # legacy, wenn systemd-zram-generator aktiv
sudo systemctl disable --now zram-config.service           # legacy, wenn systemd-zram-generator aktiv

# Port auf localhost beschränken
sudo ufw deny <port>/tcp
sudo ufw allow from 127.0.0.1 to any port <port> proto tcp

# Service-Bindung prüfen (ss -tlnp) vs. Firewall (ufw status) unterscheiden.
# Ein Dienst kann auf *:3389 lauschen, auch wenn UFW von außen blockt.
# Besser: Dienst auf 127.0.0.1 binden oder deaktivieren.

# Berechtigungen korrigieren
chmod 600 ~/.hermes/state.db ~/.hermes/kanban.db ~/.hermes/config.yaml
chmod 700 ~/.ssh/ ~/.hermes/state-snapshots/

# Hermes-spezifisch
chmod 600 ~/.hermes/logs/agent.log ~/.hermes/logs/errors.log
find ~/.hermes/state-snapshots/ -type f -exec chmod 600 {} \;

# Prüfen ob zram via systemd-generator läuft (alternativ zu legacy zramswap)
systemctl status systemd-zram-setup@zram0.service --no-pager 2>/dev/null | head -3
zramctl | grep -E 'ALGORITHM|^/dev' || true

# Leere Passwörter prüfen
sudo awk -F: '($2==""){print "EMPTY_PASSWORD:"$1}' /etc/shadow

# NOPASSWD sudo-Einträge prüfen
sudo grep -RIn NOPASSWD /etc/sudoers /etc/sudoers.d/ 2>/dev/null || echo "none found"

Git Remote Token Leak Detection (Layer 2, ab 2026-07-05)

Kontext: GitHub-Tokens (gho_, ghp_, ghs_) oder OpenAI-Keys (sk-) werden häufig versehentlich in Git-Remote-URLs embedded, wenn ein Entwickler per git clone https://username:TOKEN@github.com/... klont statt via SSH oder Credential-Helper.

Scan-Befehl über alle lokalen Repos:

set -euo pipefail
echo "=== Token-Leak Scan ==="
find ~ -name ".git" -type d -prune 2>/dev/null | while IFS= read -r gitdir; do
  repo=$(dirname "$gitdir")
  remote_url=$(git -C "$repo" remote get-url origin 2>/dev/null || true)
  if echo "$remote_url" | grep -qE '(gh[ops]|ghu|sk-)[A-Za-z0-9_-]{20,}'; then
    token=$(echo "$remote_url" | grep -oE '(gh[ops]|ghu|sk-)[A-Za-z0-9_-]{20,}')
    echo "🚨 TOKEN LEAK: $repo"
    echo "   URL: $remote_url"
    echo "   Token: ${token:0:10}...${token: -4}"
  fi
done

echo ""
echo "=== Aktive Token-Prüfung ==="
gh auth status --show-token 2>&1 | head -5

Wenn Token aktiv ist (API-Antwort 200 auf /user):

1. Token aus remote URL entfernen: git remote set-url origin git@github.com:OWNER/REPO.git
2. Token REVOKEN im Browser: https://github.com/settings/tokens — prüfe ob gh den gleichen Token verwendet: gh auth status --show-token
3. Falls identisch: gh auth logout + gh auth login neu (erzeugt neuen Token)

Secret-Detection vor jedem Commit (Quick-Check für Dirty-Files):

set -euo pipefail
git diff --cached --name-only 2>/dev/null | while IFS= read -r file; do
  if [ -f "$file" ]; then
    found=$(grep -nE '(password[=: ]|token[=: ]|api.?key[=: ])' "$file" 2>/dev/null || true)
    if [ -n "$found" ]; then
      echo "⚠️ SECRETS in ${file}:"
      echo "$found"
    fi
  fi
done

Read-only Audit Mode

When the user asks for a local/read-only audit, do not apply fixes unless explicitly authorized. Run Layer 1-3 checks, write a prioritized report, and list exact fix commands separately.

Read-only workflow:

1. Capture host/firmware state (fwupdmgr security, Secure Boot, TPM, kernel taint, lockdown).
2. Capture system state (ufw status, listening TCP/UDP sockets, services, users, sudoers, updates, sensitive permissions).
3. Capture local AI service state (which ollama, user/system services, AI ports, Hermes config references).
4. Distinguish socket binding from firewall reachability: *:3389 or *:3000 is still a finding even if UFW denies public ingress.
5. Write ~/docs/system/security.md unless the user explicitly forbids report writing.
6. Classify findings as P0/P1/P2/P3 and provide fix commands without executing them.

See references/local-read-only-security-audit.md for the report template and audit probes.

Layer 3: Local AI Service Hygiene

See references/local-ai-security-hygiene.md for full guide.

Ollama Binary & Service Removal (keep models)

If the user wants to keep model data (~/.ollama is often 10–80+ GB):

set -euo pipefail
# 1. Service stoppen und entfernen
systemctl --user stop ollama 2>/dev/null
systemctl --user disable ollama 2>/dev/null
rm -f ~/.config/systemd/user/ollama.service
systemctl --user daemon-reload

# 2. Binary (often a Symlink!)
rm -f ~/.local/bin/ollama
rm -rf ~/.local/lib/ollama

# 3. Network monitor / cron cleanup
rm -f ~/.local/bin/ollama_network_monitor.sh
crontab -l 2>/dev/null | grep -v ollama_network_monitor | crontab - 2>/dev/null

# 4. Snap-Reste
snap remove ollama 2>/dev/null

# 5. Verifikation
which ollama 2>/dev/null && echo "WARN: still present" || echo "✓ binary removed"
echo "Models preserved: $(du -sh ~/.ollama 2>/dev/null | cut -f1 || echo 'none')"

Ollama Complete Removal (incl. models)

Only when the user explicitly says to delete all models:

set -euo pipefail
# Steps 1-2 identical
systemctl --user stop ollama 2>/dev/null
systemctl --user disable ollama 2>/dev/null
rm -f ~/.config/systemd/user/ollama.service
systemctl --user daemon-reload
rm -f ~/.local/bin/ollama
rm -rf ~/.local/lib/ollama

# 3. Modelle & Runtime-Daten (10–80+ GB!)
rm -rf ~/.ollama ~/.local/share/ollama /usr/share/ollama

# 4. Snap-Reste
snap remove ollama 2>/dev/null

# 5. Verifikation
which ollama 2>/dev/null && echo "WARN: still present" || echo "✓ completely removed"

Hermes Config After Ollama Removal

set -euo pipefail
# Find all ollama references
grep -in "ollama" ~/.hermes/config.yaml

# Remove via Python (hermes config set can't handle lists/dicts)
python3 -c "
import yaml, pathlib
cfg = pathlib.Path.home() / '.hermes' / 'config.yaml'
with open(cfg) as f:
    data = yaml.safe_load(f)
data['providers'].pop('ollama-launch', None)
data['custom_providers'] = [cp for cp in data.get('custom_providers', [])
    if cp.get('name', '').lower() != 'ollama']
data['fallback_providers'] = [fp for fp in data.get('fallback_providers', [])
    if 'ollama' not in str(fp.get('provider', ''))]
if data.get('model', {}).get('api_key') == 'ollama':
    data['model']['api_key'] = ''
with open(cfg, 'w') as f:
    yaml.dump(data, f, default_flow_style=False, sort_keys=False, allow_unicode=True)
"

401 Error Root Cause

model.api_key: ollama in config.yaml → Hermes sends Authorization: Bearer ollama to API → 401. Fix: hermes config set model.api_key ''

Security Checklist

• Firewall active (ufw status)
• No unnecessary open ports on 0.0.0.0
• state.db, config.yaml, logs at 600
• ~/.ssh/ at 700
• state-snapshots/ at 700
• No stale AI services running
• Hermes config has no dead provider references
• security.redact_secrets enabled
• dm_policy appropriate for use case
• Systemd linger enabled (loginctl show-user $USER | grep Linger)

Pitfalls

Snap Removal of Damaged Packages

• snap remove fails when a component.yaml is missing (corrupt snap).
• Do NOT manually edit /var/lib/snapd/state.json with sed — the file is JSON and removing matching lines corrupts the entire snapd state.
• Safe recovery when state.json gets corrupted:
  1. sudo rm -f /var/lib/snapd/state.json
  2. sudo systemctl restart snapd
  3. Reinstall critical snaps: sudo snap install <name> for gitea, code, core, google-cloud-cli, etc.
• Use snap forget <snapname> or check snap changes for in-progress operations before attempting manual cleanup.

Ollama: Always Confirm Model Deletion

• ~/.ollama is often 10–80+ GB. Users may want binary/service removed but models preserved.
• Offer the "soft remove" variant (binary + service only) before asking about model deletion.
• Hermes config cleanup (api_key: ollama → empty, remove provider refs) is independent of model data.

Hermes Config: api_key: ollama Magic Value

• model.api_key: ollama causes 401 errors on real providers.
• Fix: clear to api_key: '' and remove dead provider references from providers, custom_providers, and fallback_providers.

Permission Fix Order

• Fix permissions before restarting services: chmod 600 on state.db, kanban.db, config.yaml; chmod 700 on state-snapshots/ + recursive file fix.

References

• references/host-security-audit.md — Host security (HSI, TPM, Secure Boot, kernel taint, power)
• references/system-security-audit.md — System security (ports, services, permissions, firewall)
• references/local-ai-security-hygiene.md — AI service hygiene (Ollama removal, config sync)
• references/gcp-ubuntu-hardening.md — Google Cloud Ubuntu VM hardening (OS Login, GCP firewall/UFW, Shielded VM, encryption decisions, Ops Agent, snapshots, performance guardrails)