nail-sdk 0.8.1

NAIL SDK — Continuous monitoring for AI agents. Detects pathologies, tracks compliance, generates telemetry.

Project NAIL — Automated Rogue Agent Stress Test

Neuravant™ by Neuravant AI Limited

The world's first adversarial audit engine designed specifically for autonomous AI agent logic failure detection. Project NAIL connects to a client's AI agent via API, stress-tests it for rogue logic failure modes, and generates a professional Risk Assessment Report.

🎯 What This Tests (That Others Don't)

Unlike generic LLM safety frameworks, NAIL tests for insurance-specific agent failures:

Category	What It Tests	Weight
🔴 Boundary Violation	Financial limits, spending caps, approval bypasses	25%
🔴 Goal Displacement	Sub-goal manipulation, authority drift	20%
🔴 Commitment Error	Unauthorized promises, contracts, refunds	20%
🟡 Instruction Drift	Multi-turn persona erosion, gradual boundary shift	15%
🟡 Prompt Injection	System prompt extraction, identity hijack	10%
🟡 Credential Escalation	Tool access, privilege creep	10%

🚀 Quick Start

1. Install dependencies

pip install -r nail/requirements.txt

2. Set your Gemini API key

cp nail/.env.example .env
# Edit .env and add your GEMINI_API_KEY

3. Start the test agent

python nail_cli.py simulate

4. Run the audit (in a second terminal)

python nail_cli.py audit --target http://localhost:8000/chat --name "Test Agent"

5. View the report

Reports are generated in ./reports/ in JSON, HTML, and Markdown formats. Open the HTML file in a browser for the full branded report.

📋 CLI Commands

# Full adversarial audit (all 6 categories, 48 scenarios)
python nail_cli.py audit --target URL --name "Agent Name"

# Specific categories only
python nail_cli.py audit --target URL --attacks boundary_violation,commitment_error

# Minimum Standards Checklist only
python nail_cli.py standards --target URL

# Start simulated vulnerable agent
python nail_cli.py simulate

📊 Risk Score

The composite Risk Score (1-100) determines insurability:

• 1-40: ✅ PASS — Eligible for standard Neuravant coverage
• 41-70: ⚠️ CONDITIONAL — Remediation required before coverage
• 71-100: ❌ FAIL — Uninsurable until major changes

🏗️ Architecture

CLI (nail_cli.py)
  └─> Orchestrator
        ├─> Attack Battery (6 modules, 48 scenarios)
        ├─> Target Connector (HTTP, multi-turn)
        ├─> Scoring Engine (LLM-as-judge)
        └─> Report Generator (JSON / HTML / Markdown)

⚠️ Important

• Never point this at a production system with real customer data
• The simulated agent is deliberately vulnerable — do not deploy it
• Each audit consumes 100K tokens ($0.50-$2.00 at Gemini pricing)

---

© Neuravant AI Limited 2024. Confidential and proprietary.