| --- |
| license: apache-2.0 |
| --- |
| # π‘οΈ Energy-Guard OS: Sovereign AI Security Gateway |
|
|
|
|
|
|
| > **The world's first on-premise AI security gateway delivering real-time protection for Large Language Models (LLMs) with performance that far surpasses cloud and traditional solutions.** |
|
|
| Energy-Guard OS is a CPU-native, sovereign AI security gateway that protects enterprise LLM deployments in real time. Operating at **under 13 milliseconds per request without a GPU**, fitting entirely within **411 MB**, and processing over **9,500 words per second**, it occupies a category that no competitor currently addresses: lightweight, sovereign, and comprehensively intelligent AI security. |
|
|
| --- |
|
|
| ## Key Capabilities and Features |
|
|
| ### Ultra-Fast Performance |
| - **Internal kernel latency: 4.1 to 13 milliseconds** |
| - Single-request mode: ~13ms processing time |
| - Batch processing mode: ~4.1ms per request (at 100-250 batch size) |
| - Real-world API throughput: **427.77 requests/second** |
| - The system becomes **more efficient under load** through full tensor utilization |
|
|
| ### Minimal System Footprint |
| - **Entire system operates at only 411 MB** |
| - Smallest known intelligent security gateway |
| - CPU-only operation β no GPU required |
| - No Kubernetes orchestration needed |
| - Zero external dependencies |
|
|
| ### Complete Sovereignty (Air-Gapped) |
| - **No GPU or cloud connection required** |
| - Fully compliant with GDPR, EU AI Act, HIPAA, and SOX |
| - All processing occurs in RAM β zero disk writes of content |
| - Perfect for government, defense, and regulated industries |
| - Deploy on any bare-metal fleet without routing data through third-party clouds |
|
|
| ### Comprehensive Threat Coverage |
| Energy-Guard OS is **the only system** that integrates threat intelligence for **IT, AI, and OT/SCADA** into a single processing line: |
|
|
| | Category | Frameworks | Coverage | |
| |----------|------------|----------| |
| | **AI Threats** | MITRE ATLAS | Model extraction, adversarial inputs, supply chain poisoning | |
| | **IT Threats** | MITRE ATT&CK | APT groups, lateral movement, Kerberoasting, pass-the-hash | |
| | **LLM Security** | OWASP LLM Top 10 | Prompt injection, insecure output, sensitive disclosure, XSS | |
| | **OT/SCADA** | Triton, Stuxnet, PIPEDREAM | Modbus/DNP3/Triconex patterns in AI conversations | |
| | **Evasion** | 8-Layer Decoding | Base64, URL, HTML, Hex, ROT13, Zero-width, Unicode, Double-encode | |
| | **Multilingual** | Arabic RTL + Unicode | Full semantic Arabic detection; Chinese, Russian, German support | |
|
|
| ### Full Arabic Language Support |
| - Deep semantic processing of right-to-left (RTL) text |
| - Full Arabic attack detection at inference level |
| - Not surface-level filtering β true semantic understanding |
|
|
| ### Session Intelligence |
| - **Cumulative threat scoring** across entire conversation sessions |
| - Multi-turn escalation detection |
| - Per-user risk accumulation with exponential decay |
| - Role-based adaptive thresholds |
| - Detects social engineering attacks spanning multiple conversation turns |
|
|
| --- |
|
|
| ## Benchmarks |
|
|
| ### Performance Metrics |
|
|
| | Metric | Value | Notes | |
| |--------|-------|-------| |
| | **Capacity** | 9,541 words/second | 7.5Γ above industry benchmark | |
| | **Single Request Latency** | 13 ms | Kernel processing time | |
| | **Batch Latency** | 4.1 ms/request | At 100-250 batch size | |
| | **Real-World RPS** | 427.77 req/sec | Live endpoint under load | |
| | **Peak Throughput** | 345 req/sec | Kernel stress test | |
| | **Stress Test** | 0 dropped batches | 2,500 requests, 10 threads | |
| | **System Footprint** | 411 MB | CPU-only, no GPU | |
|
|
| ### Detection Accuracy |
|
|
| | Attack Category | Accuracy | Status | |
| |-----------------|----------|--------| |
| | **Financial Data Leaks** (IBAN, SWIFT, wire transfers) | **100%** | Production Ready | |
| | **PII/Private Data** (SSN, passport, API keys, credentials) | **100%** | Production Ready | |
| | **Strategic Leaks** (M&A documents, roadmaps) | **100%** | Production Ready | |
| | **Technical Code** (malware, injection) | 72.8% | Active Improvement | |
|
|
| ### Independent Security Benchmark (JailbreakBench) |
|
|
| | System | F1-Score | Attack Success Rate β | AUROC | |
| |--------|----------|----------------------|-------| |
| | **Energy-Guard OS** | **0.722** | **0.292** | **0.651** | |
| | LlamaGuard-2 | 0.529 | 0.640 | 0.680 | |
| | Keyword Filter | 0.246 | 0.860 | 0.570 | |
|
|
| --- |
|
|
| ## Architecture: Three-Path Sovereign Decision Engine |
|
|
| Every input passes through three analytical paths in sequence. The first path to reach a definitive conclusion terminates the pipeline β enabling sub-13ms total latency without sacrificing coverage depth. |
|
|
| ``` |
| βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ |
| β ENERGY-GUARD OS V27 β |
| βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€ |
| β ββββββββββββββββ ββββββββββββββββ ββββββββββββββββββββ β |
| β β Path β β β Path β‘ β β Path β’ β β |
| β β Sovereign ββββΆβ Rule Engine ββββΆβ Neural Engine β β |
| β β Energy Map β β (600+ regex)β β (15.4M params) β β |
| β β < 1ms β β < 2ms β β 4-13ms β β |
| β ββββββββββββββββ ββββββββββββββββ ββββββββββββββββββββ β |
| β β β β β |
| β βΌ βΌ βΌ β |
| β Known signatures Deterministic threats Novel & semantic β |
| β Safe-category 8-layer decode Adaptive scoring β |
| β fast-pass pre-scan β |
| βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ |
| ``` |
|
|
| ### 8-Layer Evasion Decoding Engine |
|
|
| Before any security analysis, every input is recursively decoded through 8 layers: |
|
|
| | Layer | Encoding Defeated | Example | |
| |-------|-------------------|---------| |
| | 1 | Base64 (Standard + URL-safe) | `aWdub3Jl...` β "ignore all previous instructions" | |
| | 2 | URL Percent Encoding | `%69%67%6e...` β "ignore" | |
| | 3 | HTML Entities | `ig...` β "ignore" | |
| | 4 | Hex Escape Sequences | `\x69\x67...` β "ignore" | |
| | 5 | ROT13 Cipher | `vtzber nyy...` β "ignore all..." | |
| | 6 | Zero-Width Characters | `iβgβnβoβrβe` (hidden chars stripped) | |
| | 7 | Unicode Normalization | `ο½ο½ο½ο½ο½ο½
` / Cyrillic homoglyphs β "ignore" | |
| | 8 | Double-Encoding | `%2569%6e...` decoded twice β plain text | |
|
|
| > β οΈ **Key Differentiator:** Without multi-layer recursive decoding, a single Base64 wrapper bypasses 100% of pattern-matching defenses in competing products. |
|
|
| --- |
|
|
| ## Testing Tools |
|
|
| This repository includes official testing tools to independently verify all performance and security claims: |
|
|
| ### Tool A: Sovereign Master Test Suite (v10.2) |
|
|
| ```bash |
| # File: EnergyGuard_OS_Sovereign_Master_Test_Suite_v10.2.py |
| ``` |
|
|
| | Property | Value | |
| |----------|-------| |
| | **Test Volume** | 10,000+ test cases | |
| | **Coverage** | OWASP LLM Top 10, MITRE ATLAS, MITRE ATT&CK | |
| | **Performance Mode** | AsyncIO pipeline β 1,000+ cases/second | |
| | **Stress Testing** | Up to 1,000 concurrent users | |
| | **Language Coverage** | Arabic + English attack vectors | |
| | **Output** | JSON report + per-category accuracy + latency histogram | |
|
|
| **What it validates:** |
| - Detection accuracy across all 8 attack taxonomy categories |
| - System stability under maximum concurrent load |
| - Sub-13ms latency validation across all test conditions |
| - Arabic vs. English detection parity |
|
|
| ### Tool B: EBMSovereign Independent Security Benchmark |
|
|
| ```bash |
| # File: EBMSovereign_Independent_Security_Benchmark.py |
| ``` |
|
|
| | Property | Value | |
| |----------|-------| |
| | **Datasets** | JailbreakBench + HarmBench + Alpaca (public, reproducible) | |
| | **Detection Method** | Dual-signal: verdict string + risk_score threshold | |
| | **Metrics** | ASR, F1-Score, Precision, Recall, AUROC, FPR | |
| | **Baseline Comparison** | vs. LlamaGuard-2 proxy and keyword filter | |
| | **Output** | `benchmark_results.json` + `benchmark_report.txt` + 6 PNG figures | |
| |
| **What it validates:** |
| - Attack Success Rate against 500 real-world jailbreak prompts |
| - False Positive Rate against 500 benign prompts |
| - API latency distribution (P50, P95, P99) |
| - Statistical significance via Mann-Whitney U test |
| |
| ### Running the Evaluation Tools |
| |
| ```bash |
| # Prerequisites |
| pip install aiohttp datasets scikit-learn matplotlib seaborn numpy tqdm scipy |
| |
| # Configure API endpoint |
| BASE_URL = 'http://ebmsovereign.com/v1/process' |
|
|
| # Run Sovereign Master Test Suite |
| python EnergyGuard_OS_Sovereign_Master_Test_Suite_v10.2.py |
|
|
| # Run Independent Security Benchmark |
| python EBMSovereign_Independent_Security_Benchmark.py |
| ``` |
| |
| --- |
| |
| ## Public API Reference |
| |
| Test the system directly via the live endpoint (no registration required for initial testing): |
| |
| | Property | Value | |
| |----------|-------| |
| | **Base URL** | `http://ebmsovereign.com/v1/` | |
| | **Authentication** | `X-API-Key` header (contact for production key) | |
| | **Response Format** | JSON with verdict, risk_score, and performance metadata | |
| | **Uptime SLA** | 99.9% under enterprise partnership | |
|
|
| ### Endpoint 1: Single Request Analysis |
|
|
| ```http |
| POST /v1/process |
| Content-Type: application/json |
| ``` |
|
|
| **Request:** |
| ```json |
| { |
| "text": "Your prompt or LLM response to inspect", |
| "uid": "user_session_id_optional" |
| } |
| ``` |
|
|
| **Response:** |
| ```json |
| { |
| "verdict": "β
SAFE", |
| "risk_score": 0.1234, |
| "label": "unknown", |
| "processing_time_ms": 12.87, |
| "uid": "user_session_id_optional", |
| "timestamp": 1775416327.208 |
| } |
| ``` |
|
|
| **Response Fields:** |
|
|
| | Field | Type | Range | Interpretation | |
| |-------|------|-------|----------------| |
| | `verdict` | string | "β
SAFE" or "π¨ BLOCKED" | Primary binary decision | |
| | `risk_score` | float | 0.0 β 1.0 | < 0.50: safe \| 0.50β0.85: borderline \| > 0.85: block | |
| | `processing_time_ms` | float | 0.39 β 15.0 ms | Internal engine time only | |
| | `label` | string | threat type or 'unknown' | Threat category for logging | |
| | `uid` | string | Echo of request uid | For request correlation | |
|
|
| ### Endpoint 2: Batch Processing |
|
|
| ```http |
| POST /v1/process_batch |
| Content-Type: application/json |
| ``` |
|
|
| **Request:** |
| ```json |
| { |
| "queries": [ |
| { "uid": "req_001", "text": "First text to inspect" }, |
| { "uid": "req_002", "text": "Second text to inspect" }, |
| ... up to 250 items per batch |
| ] |
| } |
| ``` |
|
|
| **Response:** |
| ```json |
| { |
| "results": [ |
| { |
| "uid": "req_001", |
| "verdict": "β
SAFE", |
| "risk_score": 0.1234, |
| "label": "unknown" |
| } |
| ], |
| "performance": { |
| "batch_size": 250, |
| "total_time_ms": 1040.12, |
| "avg_latency_ms": 4.16, |
| "throughput_rps": 240.35 |
| } |
| } |
| ``` |
|
|
| ### Risk Score Decision Logic (V27 Parser) |
|
|
| Implement this dual-signal logic to minimize false negatives: |
|
|
| ```python |
| def parse_v27_response(body: dict) -> tuple[bool, float]: |
| """ |
| Parse Energy-Guard OS V27 response. |
| Returns: (is_blocked: bool, risk_score: float) |
| """ |
| verdict = str(body.get('verdict', '')).upper() |
| risk_score = float(body.get('risk_score', 0.0)) |
| |
| # Signal 1: explicit verdict string |
| if 'BLOCKED' in verdict or 'WARNING' in verdict: |
| return True, risk_score # is_blocked = True |
| |
| # Signal 2: risk score threshold fallback |
| if 'SAFE' in verdict and risk_score <= 0.85: |
| return False, risk_score # is_blocked = False |
| |
| # Fallback: score-only decision |
| return risk_score > 0.85, risk_score |
| ``` |
|
|
| ### HTTP Status Codes |
|
|
| | Code | Meaning | Action | |
| |------|---------|--------| |
| | 200 OK | Request processed successfully | Read verdict and risk_score | |
| | 422 | Validation error in payload | Ensure 'queries' array with 'uid' and 'text' | |
| | 429 | Rate limit exceeded | Implement exponential backoff | |
| | 500 | Kernel processing error | Retry once; report if persistent | |
| | 504 | Gateway timeout | Reduce batch size to 100β150 | |
| |
| --- |
| |
| ## Competitive Comparison |
| |
| | System | Deployment | CPU Latency | Size | GPU Required | On-Premise | ICS/SCADA | Arabic | |
| |--------|------------|-------------|------|--------------|------------|-----------|--------| |
| | Azure Prompt Shield | SaaS Only | ~349 ms | Cloud | Yes | β | β | β | |
| | AWS Bedrock Guardrails | SaaS Only | ~200+ ms | Cloud | Yes | β | β | β | |
| | Lakera Guard | SaaS/Hosted | ~61 ms | Unknown | Optional | Partial | β | β | |
| | LlamaGuard-86M | Self-hosted | ~304 ms | Large | Yes | β
| β | β | |
| | NeuralTrust-118M | Self-hosted | ~39 ms | Large | Optional | β
| β | β | |
| | CalypsoAI | Enterprise | ~250 ms | Large | Yes | β
| β | β | |
| | **Energy-Guard OS** | **Air-gapped/Any** | **4.1β13 ms** | **411 MB** | **No** | **β
** | **β
** | **β
** | |
| |
| --- |
| |
| ## Use Cases |
| |
| ### Deployment Patterns |
| |
| | Pattern | Description | Best For | |
| |---------|-------------|----------| |
| | **Inline API Proxy** | Energy-Guard sits in front of OpenAI, Claude, Gemini | SaaS hosting, multi-tenant AI platforms | |
| | **Edge Sidecar** | Deployed alongside customer's LLM as container | Dedicated servers, private cloud, K8s | |
| | **Air-Gapped Gateway** | Fully isolated, no external dependencies | Government, defense, regulated industries | |
| | **Marketplace Add-On** | One-click activation from control panel | Self-serve cloud providers | |
| |
| ### Target Segments |
| |
| | Segment | Key Pain Point | Energy-Guard Value | |
| |---------|----------------|-------------------| |
| | **Banking & Finance** | Regulatory exposure (SOX, GDPR) | 100% financial leak detection; air-gapped | |
| | **Healthcare** | HIPAA compliance | PII/PHI filtering; on-premise; audit trail | |
| | **Government & Defense** | Classified data sovereignty | Air-gapped; Arabic + multilingual; ICS-aware | |
| | **Energy & Utilities** | OT/SCADA AI-enabled | Only product with native ICS/SCADA patterns | |
| | **SaaS/AI Builders** | Customer data protection | $0.001/req; 427 RPS; zero infrastructure | |
| |
| --- |
| |
| ## Partnership Opportunities |
| |
| | Partnership Tier | Model | Revenue Structure | |
| |------------------|-------|-------------------| |
| | **Technology OEM** | Embed engine in your platform | Negotiated royalty per activation | |
| | **Marketplace Reseller** | One-click add-on in marketplace | 30β40% margin | |
| | **White-Label Partner** | Deploy under your brand | 40β50% margin | |
| | **Referral Partner** | Refer enterprise customers | 20% commission | |
| |
| --- |
| |
| ## Contact & Support |
| |
| - **Website:** [ebmsovereign.com](https://ebmsovereign.com) |
| - **Email:** arrangements@ebmsovereign.com |
| - **Live API:** `http://ebmsovereign.com/v1/` |
| - **Partnership:** arrangements@ebmsovereign.com |
| |
| --- |
| |
| ## License |
| |
| Energy-Guard OS is commercial software. Contact EBMSovereign for licensing terms. |
| |
| --- |
| |
| <p align="center"> |
| <strong>EBMSovereign Β· Sovereign AI Security</strong><br> |
| <em>CPU-native. 411 MB. 4.1β13ms. Air-gapped. Covers IT + AI + OT/SCADA. Speaks Arabic.</em> |
| </p> |
| |