@nanomind/runtime

NanoMind Runtime — behavioral twin for AI agent anomaly detection. The L1 layer of ARP's three-tier protection model.

Install

npm install @nanomind/runtime

Usage

import { NanoMindRuntime } from '@nanomind/runtime';

const runtime = new NanoMindRuntime('my-agent');
await runtime.initialize();

// Process behavioral events
const result = runtime.processEvent({
  agentId: 'my-agent',
  eventType: 'TOOL_CALL',
  capability: 'db:read',
  timestampDelta: 50,
  // ...
});

console.log(result.score);  // 0.0 (normal) to 1.0 (anomalous)
console.log(result.action); // 'allow' | 'alert' | 'throttle' | 'suspend' | 'kill'

How It Works

1. Baseline learning: First 100 events build a behavioral baseline
2. Anomaly scoring: 6 factors — unknown capability, timing anomaly, burst detection, L0 escalation, rare event type, error spike
3. Sub-2ms latency: No LLM — statistical model only
4. Federated learning: Gradients submitted with differential privacy

Three-Tier ARP Model

| Tier | Layer | Latency | What | |------|-------|---------|------| | L0 | Rule-based | μs | Capability enforcement | | L1 | NanoMind-Runtime | ms | Behavioral anomaly detection | | L2 | Fleet intelligence | continuous | Federated model improvement |

License

MIT