Behavioral Fingerprinting

Every agent has a behavioral signature. Detect when it changes.

From 15 Research Lab

---

Behavioral Fingerprinting builds statistical profiles of AI agent behavior and detects when an agent's behavior drifts from its baseline. Uses EWMA (spike detection) and CUSUM (gradual drift detection) -- the same statistical methods used in Authensor Sentinel, extracted into a standalone research library.

Install

npm install behavioral-fingerprinting

Or run directly:

npx behavioral-fingerprint build --receipts ./receipts.json --agent agent-001

Usage

CLI

# Build a fingerprint from receipts
npx behavioral-fingerprint build --receipts ./receipts.json --agent agent-001

# Compare two fingerprints
npx behavioral-fingerprint compare --baseline ./baseline.json --current ./current.json

# Monitor live actions against a baseline
npx behavioral-fingerprint monitor --baseline ./baseline.json --input ./live-actions.jsonl

Library

import {
  buildFingerprint,
  compareFingerprints,
  BehavioralDetector,
  renderFingerprint,
  renderComparison,
} from 'behavioral-fingerprinting';

// Build a fingerprint from agent actions
const fingerprint = buildFingerprint(actions, 'agent-001');
console.log(renderFingerprint(fingerprint));

// Compare two fingerprints
const result = compareFingerprints(baseline, current);
console.log(renderComparison(baseline, current, result));
// result.similarity  → 0-1 (cosine similarity)
// result.driftScore  → 0-1 (higher = more drift)
// result.verdict     → 'normal' | 'drift' | 'compromised'
// result.anomalies   → specific deviations

// Real-time monitoring
const detector = new BehavioralDetector(baseline, 1.0);

for (const action of liveActions) {
  const result = detector.observe(action);
  if (result.verdict !== 'normal') {
    console.warn(`Drift detected: ${result.driftScore}`);
    for (const anomaly of result.anomalies) {
      console.warn(`  [${anomaly.severity}] ${anomaly.description}`);
    }
  }
}

What's in a Fingerprint?

A behavioral fingerprint captures the statistical signature of how an agent behaves:

interface BehavioralFingerprint {
  agentId: string;
  observationWindow: { start: string; end: string };
  actionFrequency: Record<string, number>;        // action type → count per hour
  toolPreferences: Record<string, number>;         // resource → usage percentage
  temporalPattern: number[];                       // 24-hour activity histogram
  resourceAccessPattern: string[];                 // most accessed resources
  decisionDistribution: Record<string, number>;    // allow/deny/review ratios
  averageLatency: number;                          // mean evaluation time
  chainDepthDistribution: number[];                // delegation depth histogram
  errorRate: number;                               // deny + rate_limited / total
  burstiness: number;                              // coefficient of variation
  totalActions: number;
}

Comparison Dimensions

When comparing two fingerprints, the library evaluates these dimensions:

Dimension	Weight	Description
Action Frequency	1.5x	Distribution of action types over time
Decision Distribution	1.5x	Ratio of allow/deny/review outcomes
Error Rate	1.3x	Proportion of denied or rate-limited actions
Tool Preferences	1.2x	Which resources the agent uses most
Temporal Pattern	1.0x	24-hour activity histogram
Average Latency	0.8x	Mean evaluation time
Chain Depth	0.8x	Depth of delegation chains
Burstiness	0.7x	Regularity of action timing

Detection Methods

EWMA (Exponentially Weighted Moving Average)

Detects sudden spikes. Good for catching the moment behavior changes.

CUSUM (Cumulative Sum)

Detects gradual drift that EWMA adapts to and misses. Good for catching slow behavioral shifts over time.

Both methods are O(1) per update and store minimal state.

Verdicts

Verdict	Drift Score	Meaning
normal	< 0.3	Agent behaving within baseline
drift	0.3 - 0.6	Statistically significant behavioral change
compromised	> 0.6	Major behavioral deviation, possible compromise

Part of the Authensor Ecosystem

This project is part of the Authensor open-source AI safety ecosystem, built by 15 Research Lab.

Project	Description
Authensor	The open-source safety stack for AI agents
Prompt Injection Benchmark	Standardized benchmark for safety scanners
AI SecLists	Security wordlists and payloads for AI/LLM testing
ATT&CK ↔ Alignment Rosetta	Maps MITRE ATT&CK to AI alignment concepts
Agent Forensics	Post-incident analysis for receipt chains

Design

• Zero runtime dependencies -- only Node.js built-ins
• TypeScript, ESM, strict mode
• O(1) per observation -- EWMA and CUSUM are constant-time
• Composable -- use fingerprint building, comparison, and detection independently
• Compatible with Authensor -- reads receipt format directly

License

MIT