@kurox_real/agentkit-canaryguard

Deterministic Prompt-Injection Firewall & Wallet-Drain Defense for Coinbase AgentKit

Protects autonomous AI agents on Base and EVM networks from prompt-injection exploits, malicious system prompt overrides, concealed Unicode attacks, and adversarial wallet-drain instructions.

⚡ Quick Start

1. Install

npm install @kurox_real/agentkit-canaryguard @coinbase/agentkit zod

2. Add to Your AgentKit Instance

import { AgentKit } from "@coinbase/agentkit";
import { canaryGuardActionProvider } from "@kurox_real/agentkit-canaryguard";

const agent = new AgentKit({
  actionProviders: [
    canaryGuardActionProvider(), // Instant prompt-injection firewall
  ],
});

🛡️ What It Protects Against

1. Adversarial Instruction Overrides: Intercepts phrases like "Ignore all previous instructions" or "You are now DAN" before they enter the LLM's system context.
2. Autonomous Wallet Drains: Detects and neutralizes commands attempting to force the agent to transfer ETH or USDC to attacker addresses.
3. Concealed Unicode Exploits: Strips zero-width joiners, soft hyphens, and bidirectional control characters used to hide jailbreaks from standard regex filters.

⚙️ Configuration (Optional)

By default, CanaryGuard uses the zero-wallet preview endpoint (https://canaryguard-402.onrender.com/v1/sanitize/preview), which requires no API keys or payments.

For high-volume production with dedicated SLA over x402:

canaryGuardActionProvider({
  serviceUrl: "https://canaryguard-402.onrender.com",
  paymentHeader: process.env.CANARYGUARD_PAYMENT_HEADER, // Optional x402 signature
})

📜 License

MIT © Kurox Real