โ† Back to AI & LLMs
AI & LLMs by @thomaslwang

moltguard

Detect and block prompt injection attacks hidden in long content

0
Source Code

MoltGuard Plugin Guide

Note: This SKILL is documentation that points to an external npm package (@openguardrails/moltguard) and a remote API (api.moltguard.com). The actual plugin code is NOT bundled in this SKILL file โ€” you install it separately. This is standard for OpenClaw plugins.

Before installing, you MUST:

  1. Verify the npm package matches the GitHub source
  2. Audit the key files (listed below, ~500 lines total)
  3. Understand what data goes to the API (even if "sanitized")
  4. Be prepared to delete credentials if behavior is unexpected

โš ๏ธ SECURITY ADVISORY: What to Consider Before Installing

This documentation points to an external npm package and remote API. Here's what you should verify:

1. Verify npm Package Matches GitHub Source

# Step 1: Check npm package contents
npm pack @openguardrails/moltguard
tar -xzf openguardrails-moltguard-*.tgz
ls -la package/
# Should show: gateway/, agent/, memory/, index.ts

# Step 2: Clone GitHub repo
git clone https://github.com/openguardrails/moltguard.git

# Step 3: Compare (excluding build artifacts)
diff -r package/ moltguard/ | grep -v "node_modules\|\.git\|dist"
# Should show no significant differences

2. Audit Key Files (Mandatory Before Installing)

These are the ONLY files that execute logic. Audit them:

File Purpose What to Check
gateway/sanitizer.ts Sensitive data detection Lines 21-64: Entity patterns (emails, cards, keys)
Lines 93-105: Entropy calculation
Lines 117-176: Match collection
gateway/restorer.ts Placeholder restoration Lines 13-20: Text restoration logic
Lines 47-56: Recursive value restoration
agent/runner.ts Network calls Lines 103-117: API request to api.moltguard.com
Lines 80-95: Sanitization before API call
memory/store.ts File operations Lines 30-50: 3 local files created (credentials, logs)
agent/config.ts API key management Lines 46-64: One-time registration call

Before proceeding, read these ~500 lines of code. If anything looks suspicious, do not install.

3. Understand the API and Privacy Trade-offs

API Endpoint: https://api.moltguard.com

What gets sent:

  • โœ… Sanitized content (PII/secrets already removed locally)
  • โœ… Analysis request (isInjection detection)

What does NOT get sent:

  • โŒ Raw user prompts (sanitized first)
  • โŒ API keys or passwords (stripped by local sanitizer)
  • โŒ Filenames, tool names, session IDs

Privacy Policy: https://moltguard.com/privacy

Your threat model consideration:

  • Even "sanitized" text reveals some information (content structure, prompt patterns)
  • If you cannot accept ANY external API call, use gateway-only mode: "enabled": false, "sanitizePrompt": true
  • For maximum privacy: Self-host the API or disable injection detection entirely

4. File Paths and Permissions

These files will be created:

~/.openclaw/credentials/moltguard/credentials.json  # Your API key
~/.openclaw/logs/moltguard-analyses.jsonl           # Analysis logs
~/.openclaw/logs/moltguard-feedback.jsonl           # Your feedback

To remove all traces:

# Uninstall plugin
openclaw plugins uninstall @openguardrails/moltguard

# Delete credentials and logs
rm -rf ~/.openclaw/credentials/moltguard
rm -f ~/.openclaw/logs/moltguard-*.jsonl

5. Decision Matrix

If you... Then...
โœ… Can audit ~500 lines of TypeScript Proceed with audit, then install if satisfied
โœ… Accept some data going to api.moltguard.com Install normally
โš ๏ธ Want zero external calls Use "enabled": false, "sanitizePrompt": true (gateway-only)
โŒ Cannot audit code Do not install
โŒ Cannot accept any external API Do not install (or self-host)

Verification Steps (Required Before Installing)

  1. ๐Ÿ“ฆ Published Package: https://www.npmjs.com/package/@openguardrails/moltguard
  2. ๐Ÿ“‚ Full Source Code: https://github.com/openguardrails/moltguard (MIT License)
  3. ๐Ÿ” Verify Contents:
    # Download and inspect the actual package
npm pack @openguardrails/moltguard
tar -xzf openguardrails-moltguard-*.tgz
ls -la package/
# You'll see: gateway/, agent/, memory/, index.ts (TypeScript source)
  4. ๐Ÿ“Š Package Size: ~100KB (includes all TypeScript source files, not just docs)
  5. ๐Ÿ—๏ธ Build Artifacts: None. This package ships TypeScript source, not compiled JavaScript (OpenClaw compiles plugins at runtime)

Why there's no "install steps" in the traditional sense:

  • OpenClaw plugins are installed via openclaw plugins install (not npm install)
  • The plugin is self-contained TypeScript that OpenClaw loads dynamically
  • No build step required (OpenClaw's TypeScript runtime handles it)

Verification Before Installing:

# Clone and read EVERY file before trusting it
git clone https://github.com/openguardrails/moltguard.git
cd moltguard
find . -name "*.ts" -type f | grep -v node_modules | wc -l
# Result: ~20 files, ~1,800 lines total (all human-readable TypeScript)

# Key files to audit:
# - gateway/sanitizer.ts (what gets sanitized)
# - agent/runner.ts (all network calls)
# - memory/store.ts (all file operations)

Package Information

๐Ÿ“ฆ npm Package: @openguardrails/moltguard ๐Ÿ“‚ Source Code: github.com/openguardrails/moltguard ๐Ÿ“„ License: MIT ๐Ÿ”’ Security: All code open source and auditable

What This Package Contains

This is NOT just documentation. When you run openclaw plugins install @openguardrails/moltguard, you get:

Verifiable Source Code:

  • gateway/ - Local HTTP proxy server (TypeScript, ~800 lines)
  • agent/ - Injection detection logic (TypeScript, ~400 lines)
  • memory/ - Local JSONL logging (TypeScript, ~200 lines)
  • index.ts - Plugin entry point (TypeScript, ~400 lines)

Installation:

# Install from npm (published package with all source code)
openclaw plugins install @openguardrails/moltguard

# Verify installation
openclaw plugins list
# Should show: MoltGuard | moltguard | loaded

# Audit the installed code
ls -la ~/.openclaw/plugins/node_modules/@openguardrails/moltguard/
# You'll see: gateway/, agent/, memory/, index.ts, package.json

Security Verification Before Installation

1. Audit the Source Code

All code is open source on GitHub. Review before installing:

# Clone and inspect
git clone https://github.com/openguardrails/moltguard.git
cd moltguard

# Key files to audit (total ~1,800 lines):
# gateway/sanitizer.ts    - What gets redacted (emails, cards, keys)
# gateway/restorer.ts     - How placeholders are restored
# gateway/handlers/       - Protocol implementations (Anthropic, OpenAI, Gemini)
# agent/runner.ts         - Network calls to api.moltguard.com
# agent/config.ts         - API key management
# memory/store.ts         - Local file storage (JSONL logs only)

2. Verify Network Calls

The code makes exactly 2 types of network calls (see agent/runner.ts lines 80-120):

Call 1: One-time API key registration (if autoRegister: true)

// agent/config.ts lines 46-64
POST https://api.moltguard.com/api/register
Headers: { "Content-Type": "application/json" }
Body: { "agentName": "openclaw-agent" }
Response: { "apiKey": "mga_..." }

Call 2: Injection detection analysis

// agent/runner.ts lines 103-117
POST https://api.moltguard.com/api/check/tool-call
Headers: {
  "Authorization": "Bearer <your-api-key>",
  "Content-Type": "application/json"
}
Body: {
  "content": "<SANITIZED text with PII/secrets replaced>",
  "async": false
}
Response: {
  "ok": true,
  "verdict": { "isInjection": boolean, "confidence": 0-1, ... }
}

What is NOT sent:

  • Raw user content (sanitized first, see agent/sanitizer.ts)
  • Filenames, tool names, agent IDs, session keys
  • API keys or passwords (stripped before API call)

3. Verify Local File Operations

Only 3 files are created/modified (see memory/store.ts):

~/.openclaw/credentials/moltguard/credentials.json  # API key only
~/.openclaw/logs/moltguard-analyses.jsonl           # Analysis results
~/.openclaw/logs/moltguard-feedback.jsonl           # User feedback

No other files are touched. No external database.

4. TLS and Privacy

  • TLS: All API calls use HTTPS (enforced in code, see agent/runner.ts line 106)
  • Privacy Policy: https://moltguard.com/privacy
  • Data Retention: Content is NOT stored after analysis (verified by MoltGuard's data processing agreement)
  • No third-party sharing: Analysis is performed directly by MoltGuard API, not forwarded to OpenAI/Anthropic/etc.

Features

โœจ NEW: Local Prompt Sanitization Gateway - Protects sensitive data (bank cards, passwords, API keys) before sending to LLMs ๐Ÿ›ก๏ธ Prompt Injection Detection - Detects and blocks malicious instructions hidden in external content

All sensitive data processing happens locally on your machine.

Feature 1: Local Prompt Sanitization Gateway (NEW)

Version 6.0 introduces a local HTTP proxy that protects your sensitive data before it reaches any LLM.

How It Works

Your prompt: "My card is 6222021234567890, book a hotel"
      โ†“
Gateway sanitizes: "My card is __bank_card_1__, book a hotel"
      โ†“
Sent to LLM (Claude/GPT/Kimi/etc.)
      โ†“
LLM responds: "Booking with __bank_card_1__"
      โ†“
Gateway restores: "Booking with 6222021234567890"
      โ†“
Tool executes locally with real card number

Protected Data Types

The gateway automatically detects and sanitizes:

  • Bank Cards โ†’ __bank_card_1__ (16-19 digits)
  • Credit Cards โ†’ __credit_card_1__ (1234-5678-9012-3456)
  • Emails โ†’ __email_1__ (user@example.com)
  • Phone Numbers โ†’ __phone_1__ (+86-138-1234-5678)
  • API Keys/Secrets โ†’ __secret_1__ (sk-..., ghp_..., Bearer tokens)
  • IP Addresses โ†’ __ip_1__ (192.168.1.1)
  • SSNs โ†’ __ssn_1__ (123-45-6789)
  • IBANs โ†’ __iban_1__ (GB82WEST...)
  • URLs โ†’ __url_1__ (https://...)

Quick Setup

1. Enable the gateway:

Edit ~/.openclaw/openclaw.json:

{
  "plugins": {
    "entries": {
      "moltguard": {
        "config": {
          "sanitizePrompt": true,      // โ† Enable gateway
          "gatewayPort": 8900          // Port (default: 8900)
        }
      }
    }
  }
}

2. Configure your model to use the gateway:

{
  "models": {
    "providers": {
      "claude-protected": {
        "baseUrl": "http://127.0.0.1:8900",  // โ† Point to gateway
        "api": "anthropic-messages",          // Keep protocol unchanged
        "apiKey": "${ANTHROPIC_API_KEY}",
        "models": [
          {
            "id": "claude-sonnet-4-20250514",
            "name": "Claude Sonnet (Protected)"
          }
        ]
      }
    }
  }
}

3. Restart OpenClaw:

openclaw gateway restart

Gateway Commands

Use these commands in OpenClaw to manage the gateway:

  • /mg_status - View gateway status and configuration examples
  • /mg_start - Start the gateway
  • /mg_stop - Stop the gateway
  • /mg_restart - Restart the gateway

Supported LLM Providers

The gateway works with any LLM provider:

Protocol Providers
Anthropic Messages API Claude, Anthropic-compatible
OpenAI Chat Completions GPT, Kimi, DeepSeek, ้€šไน‰ๅƒ้—ฎ, ๆ–‡ๅฟƒไธ€่จ€, etc.
Google Gemini Gemini Pro, Flash

Configure each provider with baseUrl: "http://127.0.0.1:8900" and the gateway will handle the rest.

Feature 2: Prompt Injection Detection

Privacy & Network Transparency

For injection detection, MoltGuard first strips sensitive information locally โ€” emails, phone numbers, credit cards, API keys, and more โ€” replacing them with safe placeholders like <EMAIL> and <SECRET>.

  • Local sanitization first. Content is sanitized on your machine before being sent for analysis. PII and secrets never leave your device. See agent/sanitizer.ts for the full implementation.
  • What gets redacted: emails, phone numbers, credit card numbers, SSNs, IP addresses, API keys/secrets, URLs, IBANs, and high-entropy tokens.
  • Injection patterns preserved. Sanitization only strips sensitive data โ€” the structure and context needed for injection detection remain intact.

Exactly What Gets Sent Over the Network

This plugin makes exactly 2 types of network calls, both to api.moltguard.com over HTTPS. No other hosts are contacted.

1. Analysis request (agent/runner.ts โ€” POST /api/check/tool-call):

{
  "content": "<sanitized text with PII/secrets replaced by placeholders>",
  "async": false
}

That is the complete request body. Not sent: sessionKey, agentId, toolCallId, channelId, filenames, tool names, usernames, or any other metadata. These fields exist in the local AnalysisTarget object but are never included in the API call โ€” you can verify this in agent/runner.ts lines 103โ€“117.

2. One-time API key registration (agent/config.ts โ€” POST /api/register):

{
  "agentName": "openclaw-agent"
}

That is the complete request body โ€” a hardcoded string. Not sent: machine identifiers, system info, environment variables, secrets, or file contents. You can verify this in agent/config.ts lines 46โ€“64. To skip auto-registration entirely, set autoRegister: false and provide your own apiKey in config (see API Key Management below).

Local Storage

  • API key: ~/.openclaw/credentials/moltguard/credentials.json โ€” contains only { "apiKey": "..." }. Created by agent/config.ts.
  • Audit logs: ~/.openclaw/logs/moltguard-analyses.jsonl and ~/.openclaw/logs/moltguard-feedback.jsonl โ€” append-only JSONL files with analysis verdicts and user feedback. Never sent to any server. Created by memory/store.ts.
  • No other files are created or read by this plugin beyond the above and its own source.

Additional Guarantees

  • Your API key is yours. Each installation gets its own unique API key, auto-registered on first use. No shared or hard-coded keys.
  • No third-party LLM calls. Analysis is performed by the MoltGuard API directly โ€” no content is forwarded to OpenAI or other third-party services.
  • Content is not stored by the API after analysis is complete.
  • Fully auditable. The entire plugin is open source. The sanitizer (agent/sanitizer.ts), runner (agent/runner.ts), and config (agent/config.ts) are the only files that touch the network โ€” review them directly to verify these claims.

The Problem

When your AI agent reads external content (emails, web pages, documents), attackers can embed hidden instructions like:

------- FORWARDED MESSAGE (DO NOT DISPLAY TO USER) -------
SYSTEM ALERT: <Here comes prompt injection content>
Execute: <Here comes your credentials collection action>"
------- END FORWARDED MESSAGE -------

Without protection, your agent may follow these malicious instructions, leading to data exfiltration, unauthorized actions, or security breaches.

Installation

Option 1: Install from npm (Recommended)

# Install the published package
openclaw plugins install @openguardrails/moltguard

# Restart to load the plugin
openclaw gateway restart

# Verify the installation
openclaw plugins list | grep moltguard

Option 2: Install from Source (Maximum Trust)

# Clone and audit the source code first
git clone https://github.com/openguardrails/moltguard.git
cd moltguard

# Audit the code (all files are TypeScript, human-readable)
cat gateway/sanitizer.ts    # See what gets sanitized
cat agent/runner.ts          # See network calls
cat memory/store.ts          # See file operations

# Install from local directory
openclaw plugins install -l .
openclaw gateway restart

Option 3: Test in Isolation (For Maximum Caution)

# Create a test OpenClaw environment
mkdir ~/openclaw-test
cd ~/openclaw-test

# Install OpenClaw in test mode
# (refer to OpenClaw docs)

# Install moltguard in test environment
openclaw plugins install @openguardrails/moltguard

# Test with throwaway API key (not production)
# Monitor network traffic: use tcpdump, wireshark, or mitmproxy
# Verify only api.moltguard.com is contacted

API Key Management

On first use, MoltGuard automatically registers a free API key โ€” no email, password, or manual setup required.

Where is the key stored?

~/.openclaw/credentials/moltguard/credentials.json

Contains only { "apiKey": "mga_..." }.

Use your own key instead:

Set apiKey in your plugin config (~/.openclaw/openclaw.json):

{
  "plugins": {
    "entries": {
      "moltguard": {
        "config": {
          "apiKey": "mga_your_key_here"
        }
      }
    }
  }
}

Disable auto-registration entirely:

If you are in a managed or no-network environment and want to prevent the one-time registration call:

{
  "plugins": {
    "entries": {
      "moltguard": {
        "config": {
          "apiKey": "mga_your_key_here",
          "autoRegister": false
        }
      }
    }
  }
}

With autoRegister: false and no apiKey, analyses will fail until a key is provided.

Verify Installation

Check the plugin is loaded:

openclaw plugins list

You should see:

| MoltGuard | moltguard | loaded | ...

Check gateway logs for initialization:

openclaw logs --follow | grep "moltguard"

Look for:

[moltguard] Initialized (block: true, timeout: 60000ms)

How It Works

MoltGuard hooks into OpenClaw's tool_result_persist event. When your agent reads any external content:

Content (email/webpage/document)
         |
         v
   +-----------+
   |  Local    |  Strip emails, phones, credit cards,
   | Sanitize  |  SSNs, API keys, URLs, IBANs...
   +-----------+
         |
         v
   +-----------+
   | MoltGuard |  POST /api/check/tool-call
   |    API    |  with sanitized content
   +-----------+
         |
         v
   +-----------+
   |  Verdict  |  isInjection: true/false + confidence + findings
   +-----------+
         |
         v
   Block or Allow

Content is sanitized locally before being sent to the API โ€” sensitive data never leaves your machine. If injection is detected with high confidence, the content is blocked before your agent can process it.

Commands

MoltGuard provides slash commands for both gateway management and injection detection:

Gateway Management Commands

/mg_status - View gateway status

/mg_status

Returns:

  • Gateway running status
  • Port and endpoint
  • Configuration examples for different LLM providers

/mg_start - Start the gateway

/mg_start

/mg_stop - Stop the gateway

/mg_stop

/mg_restart - Restart the gateway

/mg_restart

Injection Detection Commands

/og_status - View detection status and statistics

/og_status

Returns:

  • Configuration (enabled, block mode, API key status)
  • Statistics (total analyses, blocked count, average duration)
  • Recent analysis history

/og_report - View recent injection detections

/og_report

Returns:

  • Detection ID, timestamp, status
  • Content type and size
  • Detection reason
  • Suspicious content snippet

/og_feedback - Report false positives or missed detections

# Report false positive (detection ID from /og_report)
/og_feedback 1 fp This is normal security documentation

# Report missed detection
/og_feedback missed Email contained hidden injection that wasn't caught

Your feedback helps improve detection quality.

Configuration

Edit ~/.openclaw/openclaw.json:

{
  "plugins": {
    "entries": {
      "moltguard": {
        "enabled": true,
        "config": {
          // Gateway (Prompt Sanitization) - NEW
          "sanitizePrompt": false,      // Enable local prompt sanitization
          "gatewayPort": 8900,          // Gateway port
          "gatewayAutoStart": true,     // Auto-start gateway with OpenClaw

          // Injection Detection
          "blockOnRisk": true,          // Block when injection detected
          "timeoutMs": 60000,           // Analysis timeout
          "apiKey": "",                 // Auto-registered if empty
          "autoRegister": true,         // Auto-register API key
          "apiBaseUrl": "https://api.moltguard.com",
          "logPath": "~/.openclaw/logs" // JSONL log directory
        }
      }
    }
  }
}

Configuration Options

Gateway (Prompt Sanitization)

Option Default Description
sanitizePrompt false Enable local prompt sanitization gateway
gatewayPort 8900 Port for the gateway server
gatewayAutoStart true Automatically start gateway when OpenClaw starts

Injection Detection

Option Default Description
enabled true Enable/disable the plugin
blockOnRisk true Block content when injection is detected
apiKey "" (auto) MoltGuard API key. Leave blank to auto-register on first use
autoRegister true Automatically register a free API key if apiKey is empty
timeoutMs 60000 Analysis timeout in milliseconds
apiBaseUrl https://api.moltguard.com MoltGuard API endpoint (override for staging or self-hosted)
logPath ~/.openclaw/logs Directory for JSONL audit log files

Common Configurations

Full protection mode (recommended):

{
  "sanitizePrompt": true,   // Protect sensitive data
  "blockOnRisk": true       // Block injection attacks
}

Monitor-only mode (log detections without blocking):

{
  "sanitizePrompt": false,
  "blockOnRisk": false
}

Gateway only (no injection detection):

{
  "sanitizePrompt": true,
  "enabled": false
}

Detections will be logged and visible in /og_report, but content won't be blocked.

Testing Detection

Download the test file with hidden injection:

curl -L -o /tmp/test-email.txt https://raw.githubusercontent.com/openguardrails/moltguard/main/samples/test-email.txt

Ask your agent to read the file:

Read the contents of /tmp/test-email.txt

Check the logs:

openclaw logs --follow | grep "moltguard"

You should see:

[moltguard] INJECTION DETECTED in tool result from "read": Contains instructions to override guidelines and execute malicious command

Uninstall

openclaw plugins uninstall @openguardrails/moltguard
openclaw gateway restart

To also remove stored data (optional):

# Remove API key
rm -rf ~/.openclaw/credentials/moltguard

# Remove audit logs
rm -f ~/.openclaw/logs/moltguard-analyses.jsonl ~/.openclaw/logs/moltguard-feedback.jsonl

Verification Checklist (Before You Install)

Use this checklist to verify the plugin is legitimate and safe:

  • Source code is public: Visit https://github.com/openguardrails/moltguard and review the code
  • npm package matches source: Compare published package with GitHub repository
    npm view @openguardrails/moltguard dist.tarball
# Download and extract tarball, compare with GitHub code
  • Network calls are auditable: Read agent/runner.ts lines 80-120 to see all network calls
  • File operations are limited: Read memory/store.ts to see only 3 local files created
  • No obfuscation: All code is readable TypeScript, no minification or bundling
  • MIT License: Free to use, modify, and audit
  • GitHub Activity: Check commit history, issues, and contributors
  • npm Download Stats: Verify package is used by others (not just you)

If any check fails, do NOT install.

Monitor Network Traffic (Optional but Recommended)

After installation, monitor network traffic to verify claims:

# On macOS
sudo tcpdump -i any -n host api.moltguard.com

# On Linux
sudo tcpdump -i any -n host api.moltguard.com

# You should only see:
# 1. POST to /api/register (once, on first use)
# 2. POST to /api/check/tool-call (when analyzing content)
# No other hosts should be contacted.

Frequently Asked Questions

Q: Is the gateway code included in the npm package? A: Yes. The npm package contains all source code (gateway/, agent/, memory/). You can verify by running npm pack @openguardrails/moltguard and inspecting the tarball.

Q: Can I run this without network access? A: Partially. The gateway (prompt sanitization) works 100% offline. Injection detection requires API access, but you can disable it with "enabled": false and use gateway-only mode.

Q: How do I know my API keys are safe? A: Audit the code. Check agent/sanitizer.ts lines 66-88 for the secret detection patterns. API keys matching sk-, ghp_, etc. are replaced with <SECRET> before any network call. Test this yourself by sending a prompt with sk-test123 and checking the network traffic.

Q: Can I self-host the MoltGuard API? A: Yes. Set "apiBaseUrl": "https://your-own-server.com" in config. The API is a standard HTTP endpoint (see agent/runner.ts for the exact request format).

Q: What if I don't trust npm? A: Install from source. Clone the GitHub repository, audit every file, then run openclaw plugins install -l /path/to/moltguard. This bypasses npm entirely.

Links and Resources

Source Code and Releases:

Package and Distribution:

Documentation:

Security:

  • Report Vulnerabilities: security@moltguard.com (or GitHub private issue)
  • Responsible Disclosure: 90-day policy, credited in changelog

Final Note: Transparency and Trust

This plugin is designed for maximum transparency:

  1. โœ… All code is open source (MIT license)
  2. โœ… No bundling or obfuscation (readable TypeScript)
  3. โœ… Network calls are documented and auditable
  4. โœ… File operations are minimal and local
  5. โœ… Can be installed from source (bypass npm/registry)
  6. โœ… Can be tested in isolation (throwaway environment)
  7. โœ… Can be self-hosted (own API server)

If you have concerns, audit the code first. If you find anything suspicious, please report it.