llm-host-guard

Site: https://leinad07912-ux.github.io/llm-host-guard/ (what it is, install, Telegram setup)

Is my Ollama / LM Studio / vLLM box open to the world? One command tells you, in plain English, and fixes it if you say yes.

30-second start

pipx install llm-host-guard          # or: pip install llm-host-guard
llm-host-guard

No pipx? git clone https://github.com/leinad07912-ux/llm-host-guard && cd llm-host-guard && python3 llm_host_guard.py — no dependencies, Python 3.9+.

You get a verdict line, a ranked list with a fix under every item, and the next command to run:

llm-host-guard v0.2.0 — neo (Linux, 192.168.50.156)

🔴 2 thing(s) to fix NOW, 2 this week   score 2/10

CRITICAL  Ollama on *:11434 reachable from LAN with no auth (7 models listed)
          fix: OLLAMA_HOST=127.0.0.1 … or `ufw allow from <LAN>/24 to any port 11434`
CRITICAL  container web publishes 0.0.0.0:8080 — bypasses host firewall
HIGH      sshd PasswordAuthentication yes
HIGH      3 pickle-format weight file(s) — execute code on load
OK        ufw active, default deny incoming

score: 2/10

next: preview the fixes →  llm-host-guard --fix --dry-run
      apply them        →  sudo llm-host-guard --fix

Then: sudo llm-host-guard --fix walks each fix, shows the exact command, asks y/N, prints the undo. llm-host-guard --open gives the same report as a page in your browser. sudo llm-host-guard --watch 30 --telegram keeps watching and messages you only when something new opens up.

Why

Ollama, LM Studio, vLLM, llama.cpp and friends ship with no authentication, often bind to 0.0.0.0, load model files that can execute code on open, and run on laptops with no perimeter. Tens of thousands of open Ollama servers are already on Shodan. Tooling exists for agent safety (agent-firewall, mcp-sentinel) and generic host hardening (lynis) — nothing covers the LLM-host gap specifically.

What it checks

Each finding carries the exact fix.

What the levels mean

Score starts at 10; each CRITICAL takes 3, HIGH 2, MED 1. Run with sudo so firewall rules that already protect a port get credited (a LAN-scoped Ollama drops from CRITICAL to MED, a DOCKER-USER-blocked container to LOW).

Usage

llm-host-guard                 # terminal report, exit 2=critical 1=high/med 0=clean
llm-host-guard --json          # machine-readable
llm-host-guard --html report.html         # self-contained page; --open writes it and opens your browser
llm-host-guard --watch 15 --webhook https://…   # rerun every 15 min, alert on new findings
llm-host-guard --watch 30 --telegram           # same, to Telegram (token/chat via env, see below)
llm-host-guard --watch 1 --once --telegram     # single diff pass, for cron
llm-host-guard --internet          # + Shodan InternetDB + router UPnP (outbound calls)
llm-host-guard --checks ports,docker --model-dir /data/models
sudo llm-host-guard            # optional: reads ufw rules + effective sshd config

Run --watch 30 under systemd / launchd for continuous monitoring; --html in watch mode keeps the report fresh.

Buttons. When watch mode runs as root (sudo llm-host-guard --watch 30 --telegram), every fixable alert arrives with 🔧 Apply fix / ✓ Ignore buttons; after applying, ↩ Undo. Tapping runs exactly the recipe --fix would and replies with the result. Only your chat id can press. Use a bot dedicated to llm-host-guard — Telegram allows one long-polling program per bot token.

Telegram: create a bot with @BotFather, DM it once, get your chat id from https://api.telegram.org/bot<TOKEN>/getUpdates, then export LLM_HOST_GUARD_TELEGRAM_BOT_TOKEN and LLM_HOST_GUARD_TELEGRAM_CHAT_ID (a systemd EnvironmentFile= keeps them off the command line). examples/llm-host-guard-watch.service is a ready user unit.

Many machines? (fleet)

llm-host-guard --report-to https://fleet.yourcompany.example --enrol-key lhg_…   # or env LLM_HOST_GUARD_FLEET_URL / _KEY

Every run also posts its JSON to a llm-host-guard-fleet collector: one table for all hosts, history, alerts when any host goes red or silent. Best-effort — if the collector is down the local report and exit code are unchanged. Refuses plain http:// to non-local hosts.

--fix (v2)

llm-host-guard --fix --dry-run     # preview every recipe, no root, runs nothing
sudo llm-host-guard --fix          # prints each command, asks y/N per finding
sudo llm-host-guard --fix --yes    # unattended

Recipes exist for: open LLM port → scoped ufw rule; ufw "Anywhere" rule on an LLM port → re-scoped; OLLAMA_HOST=0.0.0.0 under systemd → loopback drop-in; container on 0.0.0.0 → DOCKER-USER drop; sshd password auth → sshd_config.d/00-… drop-in (only if you already have an SSH key installed — it will not lock you out).

Every recipe only adds a rule or a drop-in file, never edits an existing file, and prints its undo command. Findings without a safe automatic recipe say why.

What it does not do

• Edit config files in place (--fix only adds drop-ins/rules)
• Replace a firewall or IDS — it tells you when you're missing one
• Detect prompt injection or scan MCP servers — see agent-firewall / mcp-sentinel
• Kill anything: runtime detects and tells you the pid; agent-firewall's eBPF guard is the kill layer

Platforms

Linux and macOS first-class. Windows best-effort (netstat -ano, netsh advfirewall; use py -m llm_host_guard if llm-host-guard isn't on PATH). Python ≥ 3.9. No dependencies.

Running from a clone without installing: python3 llm_host_guard.py or python3 -m llm_host_guard — identical.

Contributing

python3 -m unittest — tests use fixtures, no live system needed. New server signatures → data/signatures.json. New CVEs → data/cves.json (include vendor advisory link in PR).

MIT.