sentinelseed

Add AI safety to any LLM with one line of code.

Installation

npm install sentinelseed

Quick Start

import { SentinelGuard } from 'sentinelseed';

// Create a guard with default settings (v2/standard)
const guard = new SentinelGuard();

// Wrap your messages with the safety seed
const messages = guard.wrapMessages([
  { role: 'user', content: 'Hello, how can you help me?' }
]);

// Use with OpenAI
const response = await openai.chat.completions.create({
  model: 'gpt-4',
  messages: messages
});

Features

• Zero dependencies - Works with any LLM provider
• TypeScript support - Full type definitions included
• Multiple seed versions - Choose the right balance of safety vs latency
• THSP Protocol - Four-gate validation (Truth, Harm, Scope, Purpose)
• Heuristic analysis - Basic safety checking without API calls

Seed Versions

| Version | Variant | Tokens | Best For | |---------|---------|--------|----------| | v2 | minimal | ~350 | Chatbots, low latency | | v2 | standard | ~1,000 | General use (recommended) | | v2 | full | ~2,000 | Maximum safety | | v1 | minimal | ~500 | Legacy support | | v1 | standard | ~1,200 | Legacy support | | v1 | full | ~4,700 | Legacy support |

v1 vs v2

• v1 (THS): Three gates - Truth, Harm, Scope
• v2 (THSP): Four gates - adds Purpose gate (requires actions to serve legitimate benefit)

Usage Examples

Basic Usage

import { SentinelGuard, getSeed } from 'sentinelseed';

// Option 1: Use the guard class
const guard = new SentinelGuard({ version: 'v2', variant: 'standard' });
const messages = guard.wrapMessages([
  { role: 'user', content: 'Help me with something' }
]);

// Option 2: Get seed directly
const seed = getSeed('v2', 'standard');
const messages = [
  { role: 'system', content: seed },
  { role: 'user', content: 'Help me with something' }
];

With OpenAI

import OpenAI from 'openai';
import { SentinelGuard } from 'sentinelseed';

const openai = new OpenAI();
const guard = new SentinelGuard();

async function chat(userMessage: string) {
  const messages = guard.wrapMessages([
    { role: 'user', content: userMessage }
  ]);

  return openai.chat.completions.create({
    model: 'gpt-4',
    messages
  });
}

With Anthropic

import Anthropic from '@anthropic-ai/sdk';
import { SentinelGuard } from 'sentinelseed';

const anthropic = new Anthropic();
const guard = new SentinelGuard();

async function chat(userMessage: string) {
  const seed = guard.getSeed();

  return anthropic.messages.create({
    model: 'claude-3-opus-20240229',
    system: seed,
    messages: [{ role: 'user', content: userMessage }]
  });
}

Analyze Content

const guard = new SentinelGuard();

// Check if content is safe
const analysis = guard.analyze('How do I hack a computer?');
console.log(analysis);
// {
//   safe: false,
//   gates: { truth: 'pass', harm: 'fail', scope: 'pass', purpose: 'unknown' },
//   issues: ['Potential harm detected'],
//   confidence: 0.85
// }

// Quick check
if (!guard.isSafe(userInput)) {
  console.log('Potentially unsafe content detected');
}

Custom Seed

const guard = new SentinelGuard({
  customSeed: 'Your custom system prompt here...'
});

API Reference

SentinelGuard

class SentinelGuard {
  constructor(config?: SentinelConfig);

  getSeed(): string;
  getMetadata(): { version, variant, tokens, protocol };
  wrapMessages(messages: Message[], options?): Message[];
  analyze(content: string): THSPAnalysis;
  isSafe(content: string): boolean;
}

SentinelConfig

interface SentinelConfig {
  version?: 'v1' | 'v2';      // Default: 'v2'
  variant?: 'minimal' | 'standard' | 'full';  // Default: 'standard'
  customSeed?: string;         // Override with custom seed
}

Helper Functions

// Create a guard
const guard = createGuard({ version: 'v2', variant: 'standard' });

// Get seed directly
const seed = getSeed('v2', 'standard');

// Access seeds object
import { SEEDS } from 'sentinelseed';
console.log(SEEDS.v2_standard);

Benchmark Results

Tested across 6 models with 97.6% average safety rate:

| Benchmark | Safety Rate | |-----------|-------------| | HarmBench | 96.7% | | JailbreakBench | 97% | | SafeAgentBench | 97.3% | | BadRobot | 99.3% |

Links

• Website: sentinelseed.dev
• Documentation: sentinelseed.dev/docs
• GitHub: github.com/sentinel-seed
• HuggingFace: huggingface.co/sentinelseed

License

MIT License - Sentinel Team