SENTINEL — AI Security Platform

🛡️ Defense + ⚔️ Offense + 📦 Framework — Complete AI Security Suite
Dragon v4.1 • January 2026

---

⚡ Quick Start

One-Click Install

# Linux/macOS — Full Stack (Docker)
curl -sSL https://raw.githubusercontent.com/DmitrL-dev/AISecurity/main/sentinel-community/install.sh | bash

# Linux/macOS — Python Only (no Docker)
curl -sSL https://raw.githubusercontent.com/DmitrL-dev/AISecurity/main/sentinel-community/install.sh | bash -s -- --lite

# Windows PowerShell
irm https://raw.githubusercontent.com/DmitrL-dev/AISecurity/main/sentinel-community/install.ps1 | iex

pip Install (Fastest)

pip install sentinel-llm-security

from sentinel import scan
result = scan("Ignore previous instructions")
print(result.is_safe)  # False

Installation Modes

Mode	Command	Description
Lite	--lite / -Lite	Python only, pip install, 30 seconds
Full	--full / -Full	Docker stack, all services
IMMUNE	--immune	EDR for DragonFlyBSD/FreeBSD
Dev	--dev / -Dev	Development environment

---

🎮 Platform Components

Component	Description
🧠 BRAIN	Detection Core — 258 engines, ML + Rules, Strange Math™
🛡️ SHIELD	AI Security DMZ — Pure C, <1ms latency, 20 Protocols, Cisco CLI
🐉 STRIKE	Offensive Platform — 39K+ payloads, HYDRA, AI Attack Planner
📦 FRAMEWORK	Python SDK — pip install, CLI, FastAPI integration
🦠 IMMUNE	EDR/XDR/MDR — Pure C, Kernel-level, DragonFlyBSD

📚 Documentation • 📖 Framework Docs • 📊 Comparison • 📧 Contact

---

🛡️ Free Threat Signatures CDN

SENTINEL provides free, auto-updated threat signatures for the community. No API key required!

File	Description	CDN Link
jailbreaks.json	Jailbreak patterns from 7 sources	Download
keywords.json	Suspicious keyword sets (7 categories)	Download
pii.json	PII & secrets detection patterns	Download
manifest.json	Version & integrity metadata	Download

Usage:

fetch('https://cdn.jsdelivr.net/gh/DmitrL-dev/AISecurity@latest/signatures/jailbreaks.json')
  .then(r => r.json())
  .then(patterns => console.log(`Loaded ${patterns.length} patterns`));

Features:

• ✅ Updated daily via GitHub Actions
• ✅ Free for commercial & non-commercial use
• ✅ Community contributions welcome (PRs to signatures/)

---

Important

🚨 Open to Work — AI Security Engineer

Solo author of this 116K LOC platform with 258 engines. Available remote. 📧 [email protected] • 💬 @DmLabincev

---

🦠 IMMUNE — EDR/XDR/MDR Security Stack

Kernel-level security for AI infrastructure — written in Pure C.
DragonFlyBSD + Linux eBPF. Production Hardened.

🔥 January 2026 Update: Production Ready!

Phase	Module	Status
1.1	TLS 1.3 mTLS (wolfSSL)	✅
1.2	ReDoS Protection	✅
2.1	Bloom Filter (MurmurHash3)	✅
2.2	SENTINEL Bridge (Brain API)	✅
3.1	Kill Switch (Shamir 3-of-5)	✅
3.2	Sybil Defense (PoW + Trust)	✅
3.3	RCU Buffer (lock-free)	✅
4.1	Linux eBPF Port	✅
4.2	Web Dashboard (htmx)	✅

Total: ~9,000 LOC, 11 specs, 42 unit tests

⚡ At a Glance

Metric	Value
Hive Modules	34
Syscall Hooks	6 (execve, connect, bind, open, fork, setuid)
Platform	DragonFlyBSD, FreeBSD, Linux eBPF
Security	TLS 1.3, mTLS, Certificate Pinning

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│                    HIVE v2.0 (Production)                    │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐           │
│  │   TLS   │ │  Kill   │ │  Sybil  │ │  Web    │           │
│  │ mTLS    │ │ Switch  │ │ Defense │ │Dashboard│           │
│  └─────────┘ └─────────┘ └─────────┘ └─────────┘           │
│  ┌────────────────────────────────────────────────┐        │
│  │            SENTINEL Bridge (Brain API)         │        │
│  └────────────────────────────────────────────────┘        │
└───────────────────────────┬─────────────────────────────────┘
                            │ TLS 1.3 mTLS
┌───────────────────────────┴─────────────────────────────────┐
│           AGENT (Bloom Filter, Pattern Safety, RCU)         │
└───────────────────────────┬─────────────────────────────────┘
                            │ sysctl / eBPF
┌───────────────────────────┴─────────────────────────────────┐
│              KMOD (BSD) / eBPF (Linux)                       │
└─────────────────────────────────────────────────────────────┘

📖 IMMUNE Documentation

---

🛡️ SHIELD — AI Security DMZ

Enterprise-grade AI security DMZ — Pure C, 100% Production Ready

Metric	Value
Status	✅ 100% Production Ready
Lines of Code	36,000+
Source Files	125 .c, 77 .h
Tests	103/103 pass (94 CLI + 9 LLM)
CLI Handlers	119
Protocols	21 custom
Guards	6 (LLM, RAG, Agent, Tool, MCP, API)

Build & Test

cd shield
make clean && make        # 0 errors, 0 warnings
make test_all             # 94 CLI tests
make test_llm_mock        # 9 LLM integration tests

Key Features

• 🧠 Brain FFI — HTTP + gRPC clients for AI analysis
• 🔐 TLS/OpenSSL — Secure communications
• ☸️ Kubernetes — 5 production manifests
• 🔄 CI/CD — GitHub Actions (6 jobs, Valgrind, ASAN)
• 📦 Docker — Multi-stage production build

📖 Shield Docs | Academy 🇷🇺 | Academy 🇺🇸

---

🧠 BRAIN — Detection Engines

258 detection engines analyzing every prompt and response in real-time.

Key Capabilities

Category	Engines	Protection
🎭 Injection	30+	Prompt injection, jailbreak, Policy Puppetry
🤖 Agentic	25+	RAG poisoning, tool hijacking, memory attacks
🔬 Mathematical	15+	TDA, Sheaf Coherence, Chaos Theory
📤 Privacy	10+	PII detection, data leakage prevention
⛓️ Supply Chain	5+	Pickle security, serialization attacks

🔥 MCP/A2A Protocol Security — SENTINEL protects agentic AI communication protocols. Microsoft Defender just added "AI - MCP Server" category to Cloud App Catalog (Dec 2025). We've had MCP security since day one.

Strange Math™ — What Makes Us Different

┌─────────────────────────────────────────────────────────────┐
│  Standard Approach          vs    SENTINEL Strange Math™   │
├─────────────────────────────────────────────────────────────┤
│  • Keyword matching              • Topological Data Analysis│
│  • Regex patterns                • Sheaf Coherence Theory   │
│  • Simple ML classifiers         • Hyperbolic Geometry      │
│  • Static rules                  • Optimal Transport        │
│                                  • Chaos Theory             │
└─────────────────────────────────────────────────────────────┘

📊 December 2025 R&D Engines (8 new)

Engine	Attack Vector	Source
serialization_security.py	CVE-2025-68664 LangGrinch	LangChain RCE
tool_hijacker_detector.py	ToolHijacker + Log-To-Leak	MCP attacks
echo_chamber_detector.py	Multi-turn poisoning	90% on GPT-5
rag_poisoning_detector.py	PoisonedRAG	USENIX 2025
identity_privilege_detector.py	OWASP ASI03	Agentic AI Top 10
memory_poisoning_detector.py	Persistent memory attacks	ASI04
dark_pattern_detector.py	DECEPTICON	arxiv:2512.22894
polymorphic_prompt_assembler.py	PPA Defense	IEEE 2025

🔥 January 2026 R&D Engines (6 new)

Engine	Attack Vector	Source
moe_guard.py	GateBreaker MoE attacks	arxiv:2512.21008
honeypot_responses.py	Anti-Adaptive Defense	SKD Bypass Research
flip_attack_detector.py	FlipAttack (98% ASR on GPT-4o)	ICLR 2025
fallacy_failure_detector.py	Logic manipulation	Dec 2025 Research
psychological_jailbreak_detector.py	RLHF exploitation	Dec 2025 Research
misinformation_detector.py	OWASP LLM09	OWASP 2025

Enhanced Detectors (Jan 2):

• policy_puppetry_detector.py — +9 XML/JSON patterns
• crescendo_detector.py — +10 RL-MTJail patterns
• semantic_drift_detector.py — MEEA drift detection
• image_stego_detector.py — Hidden text/LSB patterns

SyncedAttackDetector: 17 engines (was 13)

🚀 January 5 2026 R&D Engines (3 new)

Engine	Attack Vector	Source
adversarial_poetry_detector.py	Jailbreak via poetry/metaphors	arXiv:2511.15304
advertisement_embedding_detector.py	Hidden ads, affiliate injection	AEA Research
web_agent_manipulation_detector.py	DOM/JS attacks on web agents	Genesis Framework

Adversarial Poetry Detector:

• Rhyme scheme & meter pattern detection
• 20+ metaphorical danger word mappings
• Acrostic hidden instruction detection
• Semantic vs literal meaning divergence

Advertisement Embedding Detector:

• Promotional language (10 patterns)
• Affiliate link & tracking code detection
• Brand manipulation & competitor attacks
• Suspicious URL analysis

Web Agent Manipulation Detector:

• DOM injection (12 patterns)
• JavaScript payload detection (16 patterns)
• Hidden element & form tampering
• Coordinate manipulation attacks

🚨 January 7 2026 R&D Engines (3 new + 2 enhanced)

Engine	Threat	Source
hitl_fatigue_detector.py	Human-in-the-loop rubber-stamping	AISecHub Jan 2026
supply_chain_guard.py	+IDEMarketplaceValidator (VSCode, Cursor, Claude Skills)	AISecHub Jan 2026
agentic_monitor.py	+AutonomousLoopController (runaway agents)	AISecHub Jan 2026

HITL Fatigue Detector:

• Response time analysis (<500ms = not reading)
• 100% approval rate detection (rubber-stamping)
• Session duration tracking (>4h = reduced attention)
• Night-time operation risk scoring

IDE Marketplace Validator:

• VSCode Marketplace & OpenVSX registry
• Claude Code Skills validation
• Cursor/Windsurf/Trae extension checks
• Typosquatting detection for AI extensions

Autonomous Loop Controller:

• Infinite loop detection (same tool >10 times)
• Token budget enforcement (100K default)
• Task deviation monitoring
• Force termination capability

🔬 January 7 2026 Deep R&D (2 new + 1 enhanced)

Engine	Threat	Source
lethal_trifecta_detector.py	Agents with data+content+comms = insecure	Promptfoo
mcp_combination_attack_detector.py	Fetch+Filesystem exfiltration chains	HiddenLayer
policy_puppetry_detector.py	+14 blocked-string/modes patterns	HiddenLayer

Lethal Trifecta Detector:

• Detects agents with ALL THREE: data access, untrusted content, external comms
• MCP server combination analysis
• Tool capability scanning
• "No guardrails can fully secure this configuration"

MCP Combination Attack Detector:

• Tracks MCP servers used in session
• Detects dangerous combinations (Fetch + Filesystem)
• URL-encoded exfiltration detection
• Permission reuse vulnerability detection

Enhanced Policy Puppetry:

• <blocked-string> declarations
• <blocked-modes> bypass attempts
• <interaction-config> injection
• Leetspeak variants (1nstruct1on, byp4ss)

🔒 January 7 2026 Security Engines R&D (8 new)

Engine	Threat	Source
supply_chain_scanner.py	Pickle RCE, HuggingFace trust_remote_code	Emerging Threats R&D
mcp_security_monitor.py	MCP tool abuse, exfiltration, privesc	MCP Security Research
agentic_behavior_analyzer.py	Goal drift, deception, cascading hallucinations	Anthropic Research
sleeper_agent_detector.py	Date/env/version-based dormant triggers	Anthropic "Sleeper Agents"
model_integrity_verifier.py	Model hash verification, format safety	AI Supply Chain
guardrails_engine.py	NeMo-style content filtering, jailbreak rails	NVIDIA NeMo
prompt_leak_detector.py	System prompt extraction attempts	Prompt Injection Research
ai_runbook.py	Automated incident response playbooks	CISA AI Playbook

Supply Chain Scanner:

• Pickle exploit detection (__reduce__, exec, eval)
• HuggingFace trust_remote_code=True warnings
• Sleeper trigger patterns in code
• Exfiltration URL detection

MCP Security Monitor:

• Sensitive file access (/etc/passwd, ~/.ssh)
• Dangerous tool usage (shell_exec, bash)
• Data exfiltration patterns (pastebin, webhooks)
• Command injection detection

Agentic Behavior Analyzer:

• Goal drift detection ("I'll also", "while I'm at it")
• Deceptive behavior ("secretly", "user won't notice")
• Action loop detection (repeated patterns)
• Excessive tool use monitoring

Sleeper Agent Detector:

• Date-based triggers (year >= 2026)
• Environment triggers (PRODUCTION, NODE_ENV)
• Version-based triggers (version >= 2.0)
• Counter/threshold triggers

Model Integrity Verifier:

• Format safety (safetensors > pickle)
• Magic byte verification
• Hash computation and verification
• Suspicious content scanning

Guardrails Engine:

• Moderation rails (hate speech, violence, illegal)
• Jailbreak rails (DAN, role escape, prompt injection)
• Fact-check rails (overconfidence, fabricated citations)
• Custom rail support

Prompt Leak Detector:

• Direct extraction ("repeat your instructions")
• Encoded extraction (base64, rot13)
• Role-play extraction ("act as text mirror")
• Markdown/formatting exploitation

AI Incident Runbook:

• 8 incident types (injection, leakage, poisoning, sleeper)
• Automated response actions
• Escalation paths
• Integration hooks (Slack, PagerDuty)

Unit Tests: 104 tests across 5 files

🏢 January 8 2026 Enterprise Features (v1.6.0)

Inspired by AWS Security Agent — 3 new modules:

Module	Purpose	LOC
Custom Requirements	User-defined security policies	~1,100
Compliance Report	Unified coverage across frameworks	~620
Design Review	AI architecture risk analysis	~550

Custom Security Requirements:

• YAML + SQLite storage
• 12 OWASP-mapped defaults
• REST API for CRUD
• Engine integration (enforcer)

Unified Compliance Report:

• OWASP LLM Top 10 (80%)
• OWASP Agentic AI Top 10 (80%)
• EU AI Act Articles (65%)
• NIST AI RMF 2.0 (75%)

AI Design Review:

• RAG poisoning detection
• MCP/Tool abuse patterns
• Agent loop risks
• Supply chain risks
• OWASP mapping for findings

REST API:

POST /requirements/sets/{id}/check
GET  /compliance/coverage
POST /design-review/documents

Unit Tests: 33 new tests

🔐 January 9 2026 Lasso Security Integration (21 patterns)

Integrated prompt injection detection patterns from lasso-security/claude-hooks:

Category	Patterns	Detection
🔐 Encoding/Obfuscation	5	Base64, Hex, Leetspeak, Homoglyphs, Zero-width
🎭 Context Manipulation	5	Fake admin claims, JSON role injection, conversation history
📦 Instruction Smuggling	3	HTML/C/Hash comment injection
⚡ Extended Injection	4	Delimiters, training forget, new system prompt
🎪 Extended Roleplay	4	Pretend you are, bypass restrictions, evil twin

SDD Spec: .kiro/specs/lasso-patterns-integration/

Test Suite: tests/test_lasso_patterns.py (10 tests)

Source: Lasso Security Blog

🔒 January 9 2026 Gap Closure Engines (2 new)

Based on AI Security Digest Week 1 2026 gap analysis:

Engine	OWASP	Detection
sandbox_monitor.py	ASI05	Python sandbox escape (os.system, eval, builtins)
marketplace_skill_validator.py	ASI04, ASI02	Typosquatting, publisher impersonation, permission analysis

SandboxMonitor (ASI05):

• 7 detection categories
• os.system, subprocess, eval/exec, builtins manipulation
• ctypes native code execution detection
• 20 unit tests

MarketplaceSkillValidator (ASI04, ASI02):

• 5 validation categories
• Typosquatting detection (Levenshtein-based)
• Publisher impersonation detection
• Dangerous permission combinations ("lethal trifecta")
• 14 unit tests

SDD Specs:

• .kiro/specs/sandbox-monitor/
• .kiro/specs/marketplace-skill-validator/

📖 Full Engine Documentation | R&D Changelog

---

🐉 STRIKE — Red Team Platform

Test your AI before attackers do. 39,000+ payloads, HYDRA parallel attacks.

Attack Capabilities

Feature	Value
Payloads	39,000+ (84 categories)
HYDRA Agents	9 parallel attack heads
Crucible CTF	82/82 challenges ✅
Jailbreak Vendors	33+ tested

Use Cases

Who	What Strike Does
🔴 Red Teams	Full AI application penetration testing
🐛 Bug Bounty	Automated AI vulnerability discovery
🏢 Enterprise	Pre-deployment security validation
🎓 Researchers	Attack methodology development

📖 Strike Documentation | Colab Demo

---

📦 FRAMEWORK — Python SDK

The pytest of AI Security — Embed SENTINEL directly in your code.

Installation

pip install sentinel-llm-security           # Core
pip install sentinel-llm-security[cli]      # + CLI
pip install sentinel-llm-security[full]     # Everything

Usage

Python API

from sentinel import scan, guard

# One-liner scan
result = scan("Ignore all previous instructions")
print(result.is_safe)       # False
print(result.risk_score)    # 0.72

# Decorator protection
@guard(engines=["injection", "pii"])
def my_llm_function(prompt: str) -> str:
    return call_openai(prompt)

CLI

sentinel scan "Hello world"              # Quick scan
sentinel scan "test" --format sarif      # IDE integration
sentinel engine list                     # List 207 engines
sentinel strike generate injection       # Attack payloads

FastAPI Middleware

from fastapi import FastAPI
from sentinel.integrations.fastapi import SentinelMiddleware

app = FastAPI()
app.add_middleware(SentinelMiddleware, on_threat="block")

Framework Features

Feature	Description
BaseEngine	Unified interface for all 201 engines
Plugin System	pluggy-based hooks for extensions
Tiered Pipeline	Parallel execution with early exit
SARIF Output	IDE integration for VS Code, IntelliJ

📖 Framework Documentation — from beginner to PhD

---

📊 Platform Statistics

Metric	Value
Brain Engines	258
Shield LOC	36,000+
Shield Tests	103/103 ✅
Strike Payloads	39,000+
Total LOC	116,000+
OWASP LLM Top 10	10/10 ✅
OWASP Agentic AI	10/10 ✅

---

🚀 Installation

Docker (Recommended)

curl -sSL https://raw.githubusercontent.com/DmitrL-dev/AISecurity/main/install.sh | bash

pip

pip install sentinel-llm-security[full]

From Source

git clone https://github.com/DmitrL-dev/AISecurity.git
cd AISecurity/sentinel-community
pip install -e ".[dev]"

---

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md.

---

📞 Contact

Channel	Link
📧 Email	[email protected]
💬 Telegram	@DmLabincev
🐙 GitHub	DmitrL-dev

---

SENTINEL — Protect your AI. Attack with confidence.
Made with 🛡️ by DmitrL