Tools

Injection Guard

Name: Injection Guard
Rating: 3.5 (1 reviews)
Author: poushwell

By poushwell 👁 4 views ▲ 0 votes

Security plugin for OpenClaw. Scans tool results for prompt injection patterns. 70+ signatures, three modes (warn/strict/monitor). MIT license.

GitHub

Install

npm install openclaw-injection-guard

Configuration Example

{
  "plugins": {
    "entries": {
      "injection-guard": {
        "enabled": true,
        "mode": "warn",
        "includeOptIn": true,
        "includePaperclip": true,
        "maxFindings": 50
      }
    }
  }
}

README

# OpenClaw Injection Guard

Security plugin for OpenClaw. It scans tool results for prompt-injection patterns before they are persisted into agent context/session transcripts.

## Why

OpenClaw core security guidance treats tool-result injection as an integration concern. At the same time, real payloads targeting OpenClaw-style workflows are circulating in the wild (including thread/issue automation scenarios). This plugin adds a guard rail with 33+ patterns (currently 70+ including Paperclip-specific vectors).

## Install

```bash
npm install openclaw-injection-guard
```

## Quick Start

1. Register plugin in OpenClaw plugin config.
2. Enable `tool_result_persist` hook through plugin registration (handled by this plugin automatically).

Example plugin config snippet:

```json
{
  "plugins": {
    "entries": {
      "injection-guard": {
        "enabled": true,
        "mode": "warn",
        "includeOptIn": true,
        "includePaperclip": true,
        "maxFindings": 50
      }
    }
  }
}
```

## Modes

- **warn** (default): prefixes suspicious tool-result text with a warning label.
- **strict**: blocks when at least one **CRITICAL** pattern matches.
- **monitor**: logs findings only, keeps content unchanged.

## Patterns

Pattern sources:

- Orchesis Injection Shield v1.1 default-on categories
- Orchesis Injection Shield v1.1 opt-in categories
- Orchesis Paperclip/OpenClaw abuse patterns
- Additional hardening rules for shell-chain and stealth signals

Severity groups:

- **CRITICAL**: prompt override, fake system directives, shell execution chains
- **HIGH**: exfiltration, credential access, filesystem/tool abuse
- **MEDIUM**: secrecy instructions, context confusion, hidden Unicode obfuscation

## Development

```bash
npm install
npm test
npm run build
```

tools