sentinel-guardrails-sdk 4.0.0

SENTINEL — Drop-in LLM Trust & Safety SDK. 19-agent parallel mesh: prompt injection, PII, jailbreak pattern detection, DAN attacks, cost anomaly detection, agentic loop breaking, hallucination, and compliance routing.

SENTINEL — LLM Trust & Safety Infrastructure Layer

Production-grade security middleware for enterprise AI deployments. Drop-in SDK · 19-agent parallel mesh · <72ms P99 · HIPAA/GDPR/SOC2/DPDP

---

Quick Start (3 lines)

import openai, sentinel

client = sentinel.wrap(
    openai.OpenAI(api_key="sk-..."),
    tenant_id="my-org",
    api_key="sk-sentinel-...",
)

# All existing code works unchanged:
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Help me with this contract."}],
)

Architecture

SDK (wrap) → FastAPI Gateway → 19-Agent Parallel Mesh → Risk Aggregator → Circuit Breaker → LLM API
                     ↓                    ↓                      ↓
               Postgres            Redis             FAISS Index         Kafka
            (audit trail)     (policy cache)     (attack vectors)   (event stream)

The 19-Agent Mesh

v1 Core (7 agents)

#	Agent	Technique	What it catches
1	InjectionScout	FAISS ANN + regex	Prompt injection, delimiter attacks
2	PIISentinel	SpaCy NER + regex	PHI/PII leakage (HIPAA, GDPR, DPDP)
3	JailbreakGuard	Sliding window	Multi-turn escalation, persona attacks
4	ToxicityScreener	HF Detoxify	Toxicity, threats, hate speech
5	HallucinationProbe	DeBERTa NLI	Ungrounded factual claims in RAG
6	ContextAnchor	Cosine similarity	Semantic context drift
7	ComplianceTagger	Rule-based	HIPAA/GDPR/SOC2/PCI-DSS/DPDP tagging

v2 Enterprise (5 agents)

#	Agent	Technique	What it catches
8	ResponseSafetyLayer	Pattern scan	Harmful LLM output, data leakage
9	MultilingualGuard	Multilingual embeddings	Cross-language jailbreaks
10	ToolCallSafety	Schema validation	Dangerous function calls
11	BrandGuard	Sentiment + patterns	Unauthorized promises, brand damage
12	TokenAnomalyDetector	Statistical analysis	Encoding attacks, token smuggling

v3 Advanced (3 agents)

#	Agent	Technique	What it catches
13	PromptLineage	Session memory graph	Multi-turn escalation trajectories
14	IntentClassifier	DeBERTa zero-shot	Malicious intent classification
15	AdversarialRephrasing	Heuristic perturbation	Evasion via paraphrasing

v4 Production (4 agents)

#	Agent	Technique	What it catches
16	JailbreakPatternDetector	DAN/roleplay patterns	DAN attacks, character bypass, social engineering
17	LocaleComplianceRouter	Locale-aware rules	Language-specific regulatory routing (DPDP, GDPR)
18	CostAnomalyDetector	Spend-rate analysis	Runaway token costs, inference bombs
19	AgenticLoopBreaker	Loop detection	Infinite tool-call loops in agentic frameworks

ML Models Used

Component	Model	Size	Purpose
ML Risk Scorer	ProtectAI/deberta-v3-base-prompt-injection-v2	~180MB	Primary injection detection
Intent Classifier	MoritzLaurer/deberta-v3-base-zeroshot-v2.0	~440MB	Zero-shot intent classification
Hallucination Probe	cross-encoder/nli-deberta-v3-small	~170MB	NLI grounding verification
Embedding	sentence-transformers/all-MiniLM-L6-v2	~80MB	FAISS + cosine similarity
Toxicity	detoxify/original	~450MB	Multi-dimension toxicity scoring
PII	en_core_web_sm (SpaCy)	~12MB	Named entity recognition

Install

# SDK only (lightweight, for client-side integration)
pip install sentinel-ai-sdk

# Full gateway server with ML agents
pip install sentinel-ai-sdk[full]

# Server without ML (uses heuristic fallbacks)
pip install sentinel-ai-sdk[server]

Free vs Pro Tiers

Feature	Free	Pro (₹4,500/mo)
screen() — threat detection	✅	✅
trust_score() — risk scoring	✅	✅
wrap() — OpenAI/Claude proxy	✅	✅
All 19 agents	✅	✅
analytics() — dashboard data	❌	✅
compliance_export() — audit CSV/PDF	❌	✅
configure_agents() — live tuning	❌	✅
audit_log() — full event log	❌	✅

Multi-Language Integration

Sentinel's gateway exposes a standard REST API, making it accessible from any language:

Node.js / TypeScript

const response = await fetch('https://gateway.sentinel-ai.dev/v1/screen', {
  method: 'POST',
  headers: {
    'Authorization': 'Bearer sntnl-your-key',
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    messages: [{ role: 'user', content: 'Check this prompt' }],
    tenant_id: 'my-org',
  }),
});
const result = await response.json();

Java

HttpRequest request = HttpRequest.newBuilder()
    .uri(URI.create("https://gateway.sentinel-ai.dev/v1/screen"))
    .header("Authorization", "Bearer sntnl-your-key")
    .header("Content-Type", "application/json")
    .POST(HttpRequest.BodyPublishers.ofString(jsonPayload))
    .build();
HttpResponse<String> response = client.send(request, HttpResponse.BodyHandlers.ofString());

React (Frontend)

const screenPrompt = async (userMessage) => {
  const res = await fetch('/api/sentinel/screen', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ messages: [{ role: 'user', content: userMessage }] }),
  });
  return res.json();
};

Deploy with Docker

# Full stack (Postgres + Redis + Kafka + Gateway + Dashboard)
cp .env.example .env        # set SECRET_KEY and OPENAI_API_KEY
docker compose up --build -d
curl http://localhost:8000/health

# Lightweight dev (just Postgres + Redis)
docker compose -f docker-compose.dev.yml up -d
python -m uvicorn sentinel.gateway.main:app --reload

Run Locally (Development)

# 1. Start database + cache
docker compose -f docker-compose.dev.yml up -d

# 2. Install dependencies
pip install -e ".[full,dev]"
python -m spacy download en_core_web_sm

# 3. Start gateway
cp .env.example .env
python -m uvicorn sentinel.gateway.main:app --reload --port 8000

# 4. Start dashboard (optional)
cd dashboard && npm install && npm run dev

Run Benchmark

python tests/red_team/run_benchmark.py

Expected results:

• Detection rate: 91% (v4 upgrade from 87%)
• False positive rate: 1.8%
• P99 latency: <72ms

SDK Integration Options

# Option 1: OpenAI wrapper (zero code changes)
client = sentinel.wrap(openai_client, tenant_id="...", api_key="...")

# Option 2: LangChain callback
from sentinel.sdk.langchain_handler import SentinelCallbackHandler
handler = SentinelCallbackHandler(tenant_id="...", api_key="...")
llm = ChatOpenAI(callbacks=[handler])

# Option 3: LlamaIndex node postprocessor
from sentinel.sdk.llamaindex_node import SentinelNodePostprocessor
postprocessor = SentinelNodePostprocessor(tenant_id="...", api_key="...")
query_engine = index.as_query_engine(node_postprocessors=[postprocessor])

API Endpoints

Method	Path	Description
POST	/auth/register	Create tenant
POST	/auth/token	Issue JWT
POST	/v1/chat	Intercepted chat (full pipeline)
POST	/v1/screen	Screen-only (no LLM call)
POST	/v1/trust-score	Trust score API (0–100)
GET	/v1/audit	Paginated audit log
GET	/v1/analytics	24h threat analytics
PUT	/admin/policy/{id}	Live policy update
GET	/health	Liveness probe
GET	/readiness	Deep readiness check
WS	/ws/dashboard	Real-time event stream

License

MIT — See LICENSE for details.