hashgraph-online/hol-guard

Open-source antivirus for AI agents: block risky tools, secret access, prompt injection, malicious packages, MCP servers, plugins, and skills at runtime.

What it solves

HOL Guard acts as a local-first security layer (similar to an antivirus) for AI agents. It prevents AI agents from performing risky actions—such as leaking secrets, executing destructive shell commands, or installing malicious packages—that could compromise a user's machine.

How it works

Guard integrates with various AI agents (like Claude Code, Cursor, and GitHub Copilot CLI) using native hooks, managed proxies, or launch overlays. It monitors runtime events, including tool calls, prompt intent, and file access. When a potentially dangerous action is detected, Guard can either allow it, block it, or pause it for human approval via a local dashboard or terminal prompt. It also provides a plugin-scanner tool for CI/CD pipelines to verify plugins and MCP servers before they are deployed.

Who it’s for

Developers and users of AI agents who want to ensure their agents operate within safe boundaries without needing a full sandbox or a cloud-managed security suite.

Highlights

  • Multi-Agent Support: Integrates with a wide range of agents including Codex, Claude Code, Cursor, Gemini CLI, and Grok.
  • Runtime Protection: Screens for prompt injection, secret exfiltration, and unsafe shell commands in real-time.
  • Human-in-the-loop: Routes ambiguous actions to an approval center for manual verification.
  • Supply Chain Security: Scans AI plugins, skills, and MCP servers for malicious content.
  • Local-First: Operates locally by default, with an optional cloud connection for team policies and synchronized evidence.
  • Security Evidence: Maintains detailed receipts of all decisions for later review and auditing.

Related

  • Project
    hardenrun/aif

    A local safety system for coding agents that prevents the execution of dangerous commands and blocks secret leaks through pre-execution blocking.

  • 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
    PrismorSec/prismor

    A runtime security framework for AI agents and coding assistants that prevents prompt injection, secret leaks, and destructive actions through policy-based enforcement.

  • Project
    duncatzat/vigils

    A local-first security control plane for AI agents that provides secret redaction, tool firewalls, and tamper-evident auditing to prevent data leaks and unauthorized actions.

  • Project
    GoPlusSecurity/agentguard

    AgentGuard is an npm‑distributed security layer for AI coding agents (Claude Code, OpenClaw, Hermes, etc.). It installs once, hooks into the agent runtime, and blocks destructive commands, protects secret files, and logs which skill triggered each action. It also offers static‑analysis scans for new skills, a daily “patrol” health‑check (8 built‑in checks), and optional cloud‑connected threat‑intel. Installation is `npm i -g @goplus/agentguard && agentguard init`; usage includes `agentguard scan`, `agentguard protect`, `agentguard patrol`, and `agentguard checkup`.