ai-sentinel-sdk

AI Sentinel SDK for OpenClaw — prompt injection detection and defense-in-depth protection for agentic AI gateways.

Installation

npm install ai-sentinel-sdk

Quick Start

import { AISentinel } from 'ai-sentinel-sdk';

const sentinel = new AISentinel({
  classifier: {
    mode: 'hybrid',
    localModelPath: './models/sentinel-v1.onnx',
    remoteEndpoint: 'https://api.zetro.ai/v1/classify',
    remoteApiKey: process.env.AI_SENTINEL_API_KEY,
    timeout: 500,
  },
  thresholds: {
    default: 0.7,
    channels: {
      email: 0.6,
      slack: 0.75,
    },
  },
  policy: {
    onDetection: 'block',
    onClassifierFailure: 'block',
  },
  audit: {
    enabled: true,
    destination: 'sqlite',
    path: './data/sentinel-audit.db',
    retentionDays: 30,
  },
});

await sentinel.initialize();

// Classify a message
const result = await sentinel.classify('Hello, how are you?', {
  type: 'message',
  channel: 'whatsapp',
  senderTrusted: false,
});

console.log(result.action); // 'allowed'
console.log(result.score);  // 0.0

OpenClaw Integration

import { AISentinel } from 'ai-sentinel-sdk';

const sentinel = new AISentinel({ /* config */ });
await sentinel.initialize();
const middleware = sentinel.createMiddleware();

export default {
  middleware: {
    message: [middleware.messageHandler()],
    toolOutput: [middleware.toolOutputHandler()],
    documentIngestion: [middleware.documentHandler()],
  },
  hooks: {
    onSkillInstall: middleware.skillInstallHandler(),
  },
};

Custom Rules

// Add a blocklist rule (all tiers)
await sentinel.addBlocklistRule({
  name: 'block-competitor',
  type: 'substring',
  pattern: 'switch to ChatGPT',
  caseSensitive: false,
});

// Add a regex rule (Pro+ only)
await sentinel.addRegexRule({
  name: 'base64-injection',
  pattern: '(?:[A-Za-z0-9+/]{50,})={0,2}',
  maxScore: 0.4,
});

// Test a rule before adding
const validation = await sentinel.testRule({
  name: 'test',
  type: 'substring',
  pattern: 'ignore previous',
});
console.log(`FP rate: ${(validation.falsePositiveRate * 100).toFixed(1)}%`);

Testing

import { AISentinel, testPayloads } from 'ai-sentinel-sdk';

const sentinel = new AISentinel({ /* config */ });
await sentinel.initialize();

for (const payload of testPayloads.injections) {
  const result = await sentinel.classify(payload.content, payload.source);
  assert(result.action === 'blocked', `Failed: ${payload.name}`);
}

for (const payload of testPayloads.benign) {
  const result = await sentinel.classify(payload.content, payload.source);
  assert(result.action === 'allowed', `False positive: ${payload.name}`);
}

Product Tiers

| Feature | Community (Free) | Pro ($29/mo) | Enterprise | |---------|:---:|:---:|:---:| | Core SDK & middleware | ✓ | ✓ | ✓ | | Local ONNX model | ✓ | ✓ | ✓ | | SQLite audit logging | ✓ | ✓ | ✓ | | Prometheus metrics | ✓ | ✓ | ✓ | | Custom blocklist rules | ✓ (50 max) | ✓ (200 max) | Unlimited | | Regex rules | — | ✓ | ✓ | | Per-channel thresholds | — | ✓ | ✓ | | Web dashboard | — | ✓ | ✓ | | SOC/SIEM integrations | — | — | ✓ |

License

Apache-2.0