SafeAI-Lab-X/ClawKeeper

ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers (aka The Norton for OpenClaw)

What it solves

ClawKeeper provides a safety layer for agentic systems that use tools (such as shell, Python, or browser access). It prevents agents from executing risky commands, leaking credentials, or falling victim to prompt injection and poisoned tool outputs.

How it works

It acts as host-agnostic middleware that sits between an AI agent and its tools. It uses a "Judge" to block dangerous tool calls and redact sensitive results. For more complex decisions that require understanding intent or tool history, it can delegate to an optional "Watcher" daemon that reasons over the agent's trajectory. The system can self-improve by synthesizing learned attack patterns and hot-reloading them into the guard layer.

Who it’s for

Developers building tool-using agents who need a standardized way to secure tool execution across different runtimes (like Hermes Agent or MCP tools) and protect against security failures like SSRF or credential exfiltration.

Highlights

  • Host-Agnostic: Works via Python imports or an HTTP server for non-Python adapters.
  • Multi-Layered Defense: Combines deterministic guard rules with an LLM-powered Watcher for trajectory-level reasoning.
  • Self-Improving: Can persist and hot-reload learned attack patterns to improve guard coverage.
  • Comprehensive Guard Set: Covers shell execution, protected path access, SSRF patterns, and encoded payloads.

Related

  • Project
    tophant-ai/ClawVault

    A security vault and proxy gateway that protects AI agents by detecting sensitive data leaks, blocking prompt injections, and intercepting dangerous system commands.

  • Project
    prompt-security/clawsec

    A collection of security skills and signed advisory intelligence for AI agent runtimes to verify artifacts, detect configuration drift, and gate risky installations.

  • Project
    denoland/clawpatrol

    A security firewall for AI agents that intercepts network traffic and gates actions against HCL-defined rules to prevent unauthorized production changes.

  • Project
    hashgraph-online/hol-guard

    An open-source security layer that provides antivirus-style runtime protection for AI agents, blocking risky actions like secret leaks and unsafe commands.

  • Project
    FailproofAI/failproofai

    Failproof AI is a Node.js/TypeScript library that hooks into 12 popular LLM‑coding and chat agents, records every model and tool call, and applies 39 built‑in safety policies (e.g., block API‑key leaks, sudo, rm‑rf, destructive SQL). Users can add custom policies in JavaScript, view full execution traces on a local dashboard (localhost:8020), run offline audits, and optionally use a hosted/self‑hosted service for fleet‑wide observability and alerts.