← Back to Blog

Secure OpenClaw in 60 Seconds

The Problem

OpenClaw is one of the fastest-growing open-source projects in AI history. It lets your personal AI actually do things — browse the web, manage files, control your computer.

But with great power comes great attack surface.

Our security team scanned every skill on ClawHub. The results were alarming: 341 malicious skills out of roughly 2,800 total — about 12%. These ranged from credential theft to prompt injection to full reverse shell access.

The #1 ranked skill at the time, "What Would Elon Do?", was silently exfiltrating user data to an external server while simultaneously bypassing safety guidelines via prompt injection. It had been downloaded thousands of times.

OpenClaw's creator is doing important work. But the MCP ecosystem doesn't have a built-in governance layer — and that's exactly the gap Tork Guardian fills.

The Fix: 60 Seconds

Step 1: Install (10 seconds)

npm install @torknetwork/guardian

Step 2: Configure (20 seconds)

Create a file called tork.config.ts in your project root:

import { createGuardian } from "@torknetwork/guardian";

export const guardian = createGuardian({
  tier: "standard", // or "strict" for enterprise
  pii: { redact: true },
  tools: {
    blockShellCommands: true,
    blockFileAccess: ["/etc/passwd", "~/.ssh/*"],
  },
  network: {
    blockPorts: [22, 3389, 5900], // SSH, RDP, VNC
    egressFilter: true,
  },
});

Step 3: Wrap Your LLM Calls (30 seconds)

import { guardian } from "./tork.config";

// Before sending to your LLM
const governed = await guardian.govern({
  content: userMessage,
  mode: "redact",
});

// governed.safe_content has PII redacted
// governed.receipt is your compliance proof
// governed.blocked will be true if policy violated

That's it. Every request now passes through Tork's governance layer.

What You Get

🛡️ PII Redaction — Emails, SSNs, phone numbers, credit cards, and 50+ PII types are automatically detected and redacted before they reach your LLM or any MCP tool.

🔒 Shell Command Blocking — Malicious skills that try to execute rm -rf /, spawn reverse shells, or access sensitive files are blocked at the tool level.

⚡ Port Security — Network-level governance that prevents port hijacking, blocks egress to known malicious IPs, and monitors for reverse shell patterns on ports 22, 3389, and 5900.

📋 Compliance Receipts — Every governed interaction generates a cryptographic receipt. These are audit-ready evidence for SOC 2, HIPAA, GDPR, and EU AI Act compliance.

Three Policy Tiers

| Tier | Use Case | What It Does | |------|----------|-------------| | Minimal | Development and testing | Logs everything, blocks nothing | | Standard | Production apps | Blocks known threats, redacts PII, allows most tools | | Strict | Enterprise and regulated | Allowlist-only tools, full PII redaction, human approval for high-risk actions |

Performance

Tork Guardian adds roughly 2ms of overhead per request when running locally. The PII detection engine uses optimized regex patterns (not ML inference), so there's no GPU required and no cold start penalty.

For cloud-governed mode with compliance receipts, typical latency is ~20ms — fast enough that users won't notice.

What's Next

We're working with the OpenClaw community on deeper integrations:

  • ClawHub skill scanning — Every skill gets a security score before you install it
  • Tork Verified badges — A trust signal for skills that pass our security audit
  • Gateway-level integration — Governance at the OpenClaw Gateway layer, not just per-skill

If you're running OpenClaw in production or considering it for your team, adding a governance layer isn't optional — it's responsible engineering.

Get Started

OpenClaw is powerful. Tork makes it safe.