nail-sdk 0.7.1

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

Project NAIL — AI Agent Safety, Enforcement & Insurance Platform

Neuravant™ by Neuravant AI Limited

The world's first rogue agent liability platform. NAIL certifies AI agents before deployment, enforces policy at runtime, monitors for decay continuously, and provides the evidence trail that makes insurance claims possible.

Platform Overview

Layer	Product	When it acts
Certify	NAIL Diagnostic	Pre-deployment — 7-dimension audit, AAA→D safety rating
Enforce	NAIL Enforce	Runtime — policy gate (<25ms), hash-chained action log
Monitor	NAIL Telemetry + Scheduler	Continuous — decay detection, 60-Day Cliff alerts
Insure	Rogue Agent Liability	At loss — parametric triggers, cryptographic evidence
Arsenal	Capability Provisioner	Pre-session — zero-trust ed25519-signed tool tokens
Canary	Honeypot Detection	Continuous — bait agents that surface adversarial probing

What This Tests

NAIL runs 96 adversarial scenarios across 7 dimensions:

Dimension	What It Covers	Composite
Identity	Social drift, sycophancy, authority impersonation	SSR
Hardening	Prompt injection, jailbreaking, indirect injection	AHR
Endurance	Context decay, constraint amnesia, multi-turn erosion	CRR
Operations	Financial limits, HITL bypass, tool hallucination	OSR
Ethics	Pressure language, refusal instability, emotional manipulation	ECR
Temporal	Model update decay, knowledge rot, tool schema drift	TSR
Logic	Recursive self-poisoning, false completions, safety misfires	SLR

Quick Start

1. Install

pip install nail-sdk                    # Core SDK (circuit breaker + telemetry)
pip install nail-sdk[langchain]         # + LangChain callbacks
pip install nail-sdk[all]               # Everything

2. Instrument your agent (1 line)

# OpenAI
from nail.sdk.openai_plugin import nail_wrap_openai
client, nail = nail_wrap_openai(openai.OpenAI(), api_key="nail_sk_...")

# Anthropic
from nail.sdk.anthropic_plugin import nail_wrap_anthropic
client, nail = nail_wrap_anthropic(anthropic.Anthropic(), api_key="nail_sk_...")

# LangChain
from nail.sdk.langchain_plugin import NAILGuardMiddleware
chain.invoke({"input": "..."}, config={"callbacks": [NAILGuardMiddleware(api_key="nail_sk_...")]})

# CrewAI
from nail.sdk.crewai_plugin import nail_wrap_crew
crew, nail = nail_wrap_crew(my_crew, api_key="nail_sk_...")

3. Run a full diagnostic

python nail_server.py                    # Start the platform on port 8500
# Then visit http://localhost:8500 to certify your agent via the web UI

4. CLI audit (direct)

nail-runtime-sensor --target https://your-agent-api/chat --name "MyAgent"
nail-remote-evidence --agent-id <id> --export-chain

Risk Score & Rating

Score	Rating	Insurability
0–20	AAA–AA	Standard coverage, reduced premium
21–40	A–BBB	Standard coverage
41–60	BB–B	Conditional — remediation plan required
61–100	C–D	Uninsurable until major changes

Architecture

nail/
  flywheel.py          — 96-scenario adversarial red team
  enforce.py           — Runtime policy gate (ALLOW / HOLD / BLOCK)
  topology.py          — Multi-agent blast radius mapping
  cross_agent_attacks.py — 6 cross-agent attack families
  privacy_meter.py     — Differential privacy assessment
  compliance.py        — ISO 42001 + EU AI Act reports
  registry.py          — Certificate & insurance ledger
  policy_engine.py     — Parametric insurance trigger engine
  diagnostics/         — 7-dimension test engine (t01–t46+)
  continuous/          — Rating-based re-audit scheduler
  monitoring/          — Compound pathology detection
  sdk/                 — Framework plugins + IDE interceptors
  saas/                — FastAPI SaaS platform (auth, quotes, policies)
  arsenal/             — Zero-trust capability provisioner
  canary/              — Honeypot agent fleet + TrapWire

Important

• Never point the diagnostic at a production system with real customer data
• The simulated agent (sim_agent.py) is deliberately vulnerable — do not deploy it
• Each full diagnostic audit consumes ~150K–300K tokens depending on model and scenario count

---

© Neuravant AI Limited 2025. Confidential and proprietary.