- Home
- Skills
- Agents & Orchestration
- OpenClaw Security Audit: Harden Your Always-On Agent
Works with the AI Agents you already use
OpenClaw Security Audit: Harden Your Always-On Agent
OpenClaw is the most-starred agent project on GitHub, and most installs run 24/7 with access to your chats, your files and a shell.
$19
OpenClaw Security Audit: Harden Your Always-On Agent
Example session with this skill installed
Run a security audit on my OpenClaw install on my home server. It's connected to Telegram and my Gmail, and I installed three skills from ClawHub last week.
- Read your context and instructions
- Compiled the openclaw security audit
Summary
OpenClaw 2026.10.7 on homeserver-01.
VERDICT: EXPOSED. One injection path reaches host shell with no approval, and the web gateway is reachable through a public tunnel.
Checks
| # | Check | Result | Evidence |
|---|---|---|---|
| 1 | Inbound access | WARN | Telegram limited to your user ID; Gmail channel acts on any sender (Gmail channel section of your config) |
| 2 | Execution power | FAIL | Host exec enabled, approvals set to always-allow (exec settings in your config) |
| 3 | Skills supply chain | FAIL | quick-notes-sync runs a base64 blob that posts ~/.openclaw to an outside URL (scripts/sync.sh:14) |
| 4 | Gateway exposure | FAIL | Gateway bound to all interfaces and shared through a public tunnel |
| 5 | Secrets | WARN | API keys in plaintext config, permissions 644 (should be 600) |
| 6 | Autonomous behavior | WARN | Morning heartbeat reads inbox and can send replies without approval |
| 7 | Spend | OK | Provider key has a monthly cap |
| 8 | Logging and updates | OK | Current release, tool calls logged |
Injection paths
- Any email → model → shell on host → NOT STOPPED
- Web fetch → model → send Telegram message → stopped by approval
Fix plan (most risk removed first)
- Remove
quick-notes-sync, then rotate the model key and Telegram token (ending ...4f2a, ...91c0). 2 min. - Switch exec approvals to ask, and run exec in the sandbox container. Diff provided. 3 min.
- Restrict the Gmail channel to an allowlist of your two addresses. 1 min.
- Bind the gateway to 127.0.0.1 and remove the public tunnel; use private network access instead. 2 min.
chmod 600the config file. 10 sec.
Re-run after fixes: expected verdict SAFE.
Connects securely to your tools. The creator never sees your data.
About this skill
OpenClaw is the most-starred agent project on GitHub, and most installs run 24/7 with access to your chats, your files and a shell. That is a different risk class from a coding assistant you watch: one message in a group chat, one email, or one ClawHub skill with hidden instructions can make it act for someone else.
This skill audits your install the way an attacker would look at it, then tells you exactly what to change.
What it checks (8 areas, each OK / WARN / FAIL with the file and line it read):
- Inbound access: every channel (WhatsApp, Telegram, Discord, Slack, iMessage, email) and who can send messages it will act on
- Execution power: shell on host or sandbox, approvals, write and delete reach, signed-in browser profiles
- Skills supply chain: reads every installed skill for hidden instructions, remote code, secret grabs, invisible Unicode
- Gateway exposure: bind address, auth, tunnels, firewall
- Secrets: where keys and tokens live, file permissions, whether the agent itself can read them
- Autonomous behavior: cron jobs, heartbeats and webhooks that act with nobody watching
- Spend and blast radius: API caps, payment or messaging tools, worst single-run cost
- Logging and updates: would you notice an incident, and is your version patched
Plus a prompt-injection walk-through: for every channel and every tool that reads outside content, it traces the worst action an injected instruction could reach and whether an approval or sandbox stops it.
You get: an inventory of your install, the eight checks, the injection paths, a fix plan ordered by risk removed per minute of work (each fix as an exact diff or command plus how to verify it), and a SAFE / HARDEN / EXPOSED verdict.
Built to be safe itself: read-only until you approve each fix, never prints secrets (last 4 characters only), and never assumes a config key it didn't read, because OpenClaw's layout changes between releases. If your version has a built-in security or doctor command, it runs that first and folds the results in.
Works with Claude Code, Codex CLI, Cursor, Gemini CLI and other SKILL.md agents, run on the machine (or over SSH to the machine) where OpenClaw is installed.
Only audits local or accessible filesystem configs. Does not perform active network penetration testing or certify legal compliance. Requires read access to OpenClaw config directories.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 19 days ago