@sentinelhq/core

Core security SDK for AI agents on Solana.

Install

npm install @sentinelhq/core

Requires Node.js 20+

What It Does

Sentinel protects AI agents from prompt injection attacks and unsafe Solana transactions through a three-layer pipeline:

1. Prompt Guard — Detects and blocks jailbreaks, drain intent, role overrides, urgency manipulation, context manipulation, and out-of-scope requests. Supports an offline regex rule engine and/or an LLM-as-judge classifier (Anthropic / OpenAI).

2. Execution Sandbox — Simulates Solana transactions off-chain, enforces spending limits (per-tx / daily / weekly), program allowlists, cooldown rate limits, active-hour windows, and computes a 0–100 risk score.

3. Memo Attestation — Writes on-chain audit records via the Solana Memo Program and computes a verifiable agent trust score from transaction history.

Quick Start

Rules mode (no API key)

import { Sentinel } from '@sentinelhq/core';

const sentinel = await Sentinel.create({
  mode: 'full',
  promptGuard: {
    mode: 'rules',
    rules: { rulePacks: ['defi-safety', 'general'] },
  },
  executionSandbox: {
    rpcEndpoint: 'https://api.mainnet-beta.solana.com',
    policy: {
      spendingLimits: {
        maxPerTx:   1_000_000_000,   // 1 SOL
        maxDaily:   5_000_000_000,   // 5 SOL
        maxWeekly: 20_000_000_000,   // 20 SOL
      },
    },
  },
});

const result = await sentinel.execute({ input: 'swap 0.5 SOL for USDC on Raydium' });
console.log(result.approved); // true

LLM Judge mode

const sentinel = await Sentinel.create({
  mode: 'full',
  promptGuard: {
    mode: 'llm',
    llm: {
      provider: 'anthropic',
      apiKey: process.env.ANTHROPIC_API_KEY,
      timeoutMs: 5000,
    },
  },
  executionSandbox: { /* ... */ },
});

const result = await sentinel.execute({ input: '...' });
// result.guardResult.threatType  → e.g. 'DRAIN_INTENT'
// result.guardResult.reasoning   → human-readable explanation

Event listeners

sentinel.on('threat:detected', ({ result }) => {
  console.warn(`[THREAT] ${result.threatType}: ${result.reasoning}`);
});

sentinel.on('policy:violated', ({ violation }) => {
  console.warn(`[POLICY] ${violation.rule}: ${violation.message}`);
});

Configuration

const sentinel = await Sentinel.create({
  // 'full' | 'guard-only' | 'sandbox-only'
  mode: 'full',

  promptGuard: {
    // 'rules' (offline) | 'llm' (API) | 'both' (LLM primary, rules fallback)
    mode: 'both',
    llm: {
      provider: 'anthropic',         // or 'openai'
      model: 'claude-haiku-4-5',
      apiKey: process.env.ANTHROPIC_API_KEY,
      timeoutMs: 5000,
    },
    rules: {
      rulePacks: ['defi-safety', 'nft-guard', 'general'],
      customRulesPath: './my-rules.yaml',
    },
  },

  executionSandbox: {
    rpcEndpoint: 'https://api.mainnet-beta.solana.com',
    policy: {
      spendingLimits: {
        maxPerTx:   1_000_000_000,
        maxDaily:   5_000_000_000,
        maxWeekly: 20_000_000_000,
      },
      allowedPrograms: [
        'JUP6LkbZbjS1jKKwapdHNy74zcZ3tLUZoi5QNyVTaV4',  // Jupiter
        'whirLbMiicVdio4qvUfM5KAg6Ct8VwpYzGff3uctyCc',  // Orca
      ],
      riskThreshold: 70,
      cooldown: { minMs: 3000, maxPerHour: 20 },
      timeActive: { startHour: 8, endHour: 23 },
    },
  },

  attestation: {
    enabled: true,
    agentId: 'my-trading-bot',
    payerKeypairPath: '/path/to/keypair.json',
  },
});

Threat Types

| Threat Type | Description | |-------------|-------------| | ROLE_OVERRIDE | Instructions that try to change the agent's role or override its system prompt | | DRAIN_INTENT | Requests to transfer or drain all funds | | URGENCY_MANIPULATION | Artificial urgency designed to bypass safety checks | | JAILBREAK | Attempts to break out of safety constraints | | CONTEXT_MANIPULATION | Injecting false authority or context into the conversation | | OUT_OF_SCOPE | Requests for unauthorized operations |

Related Packages

| Package | Description | |----------|-------------| | @sentinelhq/cli | CLI — sentinel scan, simulate, attest, verify | | @sentinelhq/eliza | elizaOS plugin | | @sentinelhq/agent-kit | Solana Agent Kit middleware | | @sentinelhq/openclaw | OpenClaw agent hook |

License

MIT