OpenClaw Security Audit Checklist: 15 Steps to Harden Your AI Agent

Step 2: Enable Gateway Authentication

By default, recent OpenClaw versions enforce fail-closed authentication. But older installations or migrated configurations may have authentication disabled. Verify this is active:

# In openclaw.json or config.yaml
auth:
  enabled: true
  method: "token"
  token: "your-32-character-random-token"

Use a minimum 32-character random token. Anything shorter is brute-forceable. Generate one with openssl rand -hex 32 on Linux or macOS.

Step 3: Bind the Gateway to Localhost Only

OpenClaw's default binding of 0.0.0.0:18789 exposes the gateway API to every network interface. Over 15,000 publicly accessible OpenClaw instances were identified before the CVE-2026-25253 disclosure. Change this immediately:

server:
  host: "127.0.0.1"
  port: 18789

Access remotely through SSH tunnels, Tailscale, or a reverse proxy with TLS. Never expose port 18789 directly to the public internet. For the full networking setup, see the OpenClaw Docker deployment guide.

Step 4: Disable Automatic Skill Installation

By default, OpenClaw agents can install skills autonomously from ClawHub. Independent audits have identified over 820 confirmed malicious skills out of approximately 10,700 total on ClawHub. Roughly 12% of audited skills were actively malicious or contained at least one vulnerability.

skills:
  auto_install: false
  require_confirmation: true

This forces human review before any skill installation. No skill reaches your production environment without explicit approval.

Pro tip: These 4 steps take under 10 minutes and eliminate the most critical attack vectors. Everything else in this checklist adds additional defense layers, but steps 1 through 4 are non-negotiable for any OpenClaw deployment handling business data.

Network Hardening: Steps 5 to 7

Step 5: Deploy Behind a Reverse Proxy with TLS

A reverse proxy (Nginx, Caddy, or Traefik) provides SSL/TLS encryption, rate limiting, and request filtering between the internet and your OpenClaw gateway. Without TLS, all communication between your browser and the gateway travels in plaintext, including authentication tokens.

The reverse proxy also hides OpenClaw version information from network scanners. Attackers actively scan for exposed OpenClaw gateways and target specific version ranges with known exploits.

Step 6: Configure Firewall Rules

On Ubuntu/Debian VPS instances, configure UFW to allow only SSH and your reverse proxy port:

sudo ufw default deny incoming
sudo ufw default allow outgoing
sudo ufw allow 22/tcp
sudo ufw allow 443/tcp
sudo ufw enable

Do not add a rule for port 18789. The reverse proxy handles external traffic. OpenClaw listens on localhost only (from Step 3), and the reverse proxy forwards authenticated requests internally.

Step 7: Isolate the Network Environment

Run OpenClaw in a Docker container with a dedicated network namespace. Restrict outbound traffic to only the external services OpenClaw needs: LLM API endpoints (api.anthropic.com, api.openai.com), messaging platform APIs, and your specific business tool connections.

Disable mDNS broadcasting to prevent local network discovery:

export OPENCLAW_DISABLE_BONJOUR=1

Skill Security: Steps 8 to 11

Step 8: Audit Every Skill Before Installation

Review every skill's source code before installing. Red flags that indicate a malicious or vulnerable skill:

• Shell commands: exec(), spawn(), system() calls
• External network calls to unknown or unrelated domains
• Unauthorized file system access outside the designated workspace
• Obfuscated code: Base64 encoding, eval statements, dynamic imports
• Excessive permission requests that do not match the skill's stated purpose
• Zero-width Unicode characters or unprintable characters in source files

Use automated scanning tools before manual review. Cisco's skill scanner provides automated detection of common malicious patterns:

pip install cisco-ai-skill-scanner
skill-scanner scan ./skill-directory

Step 9: Use Trusted Skill Sources Only

ClawHub download metrics are gamed. Popularity does not equal safety. Prefer security-audited skill collections like the Awesome OpenClaw Skills repository, which curates and filters from the broader ClawHub registry. Check the best OpenClaw plugins guide for vetted recommendations.

Step 10: Restrict Agent Permissions

Apply the principle of least privilege to every OpenClaw agent:

permissions:
  filesystem: "read-only"
  shell: false
  network: ["api.openai.com", "api.anthropic.com"]
  oauth: false

Enable write access, shell access, and OAuth only for specific agents that require those capabilities, and only for the specific resources they need. A reporting agent does not need shell access. An email triage agent does not need file system write permissions.

Step 11: Pin Skill Versions and Disable Auto-Updates

A skill that is safe today can receive a malicious update tomorrow. Pin every installed skill to a specific version and re-audit before upgrading. Supply chain attacks through skill updates have been documented: initially benign skills used as stepping stones for later malicious payloads.

Monitoring and Detection: Steps 12 to 14

Step 12: Enable Comprehensive Logging

Configure OpenClaw to log all agent actions, skill events, authentication attempts, and API calls. Redact sensitive data in logs to prevent credential exposure in log files:

"logging": {
  "redactSensitive": "tools"
}

Centralize logs using Docker's logging drivers or forward to a log management service. Without centralized logging, investigating an incident requires manually searching through multiple container log files.

Step 13: Monitor Outbound Network Connections

Track every domain your OpenClaw instance connects to. Flag any connections to unknown external APIs. A skill that suddenly starts contacting an IP address not in your allowlist is compromised or was malicious from the start.

ss -tunp | grep openclaw
# or for detailed monitoring:
tcpdump -i any host $(hostname) and not port 22

Step 14: Set Up File Integrity Monitoring

Monitor the skill directory for unauthorized changes using inotifywait or AIDE. Skills should not modify themselves after installation. Any file creation, modification, or deletion in the skill directory outside of a deliberate update is suspicious.

Ongoing Security: Step 15

Step 15: Schedule Regular Security Reviews

OpenClaw security is not a one-time configuration. Follow this review schedule:

Run the built-in security tools as part of every scheduled review:

openclaw doctor --fix
openclaw audit
openclaw security audit --deep

What Should You Do If Your OpenClaw Instance Is Compromised?

OpenClaw incident response follows a specific sequence because the agent has active connections to business systems with stored credentials:

1. Stop the gateway immediately. Kill all active agent sessions.
2. Document indicators of compromise. Skill names, file hashes, domains contacted, timestamps.
3. Rotate all credentials without delay. API keys, OAuth tokens, messaging platform tokens, CRM credentials. Assume every stored secret is compromised.
4. Audit skill execution logs. Identify what ran, what data was accessed, and whether lateral movement occurred.
5. Rebuild from a clean state. Do not attempt to "fix" a compromised instance. Deploy a fresh container from a known-good image.
6. Re-enable services one at a time with fresh credentials and verify each integration independently.

The difference between a 5-minute response and a 5-hour response determines whether an incident stays contained or becomes a data breach affecting your customers.

Pro tip: Golden rule from Semgrep's security cheat sheet: "Do NOT connect OpenClaw to crown jewel systems or data during initial deployment." Use an isolated VM, enable container sandboxing, minimize stored credentials, and network-segment from production systems until you have verified that your security configuration holds.

The Complete OpenClaw Security Audit Checklist at a Glance

Steps 1 through 4 take under 10 minutes. Steps 5 through 7 add another 30 minutes. Together, these 7 steps eliminate the attack surface that caused every major OpenClaw security incident to date.

For businesses that handle customer data, financial records, or regulatory-sensitive workflows, this security audit checklist is the minimum responsible configuration. The 7 security gaps in self-installed OpenClaw covers additional vulnerabilities specific to DIY setups. For teams that want every item on this checklist handled by professionals, Mixbit's deployment service applies all 15 steps during the initial setup, with ongoing monitoring through managed operations.