STANDALONE PLATFORM — AUTONOMOUS DEFENCE ENGINE

Red Specter SENTINEL PRIME

177 modules watching. One mind reasoning. Zero attacks completing.

220
Tests
177
Modules Correlated
8
Attack Chains
3
Gate Logic
v1.0.0
Version
Documentation → GitHub → 4/4 Validation →

Overview

SENTINEL PRIME is a standalone autonomous defence platform that unifies the entire Red Specter AI Shield ecosystem. It continuously ingests alerts from all 177 AI Shield modules, runs multi-source correlation to identify coordinated attack chains, invokes DeepSeek R1 (deepseek-r1:7b) via local Ollama on the RTX 3090 for structured threat reasoning, and applies rigorous 3-gate logic before triggering M99 escalation.

Unlike the individual AI Shield modules which each defend a single surface, SENTINEL PRIME operates at the meta-layer — it sees across all modules simultaneously and reasons about the attack as a whole, not its constituent parts. Intelligence feeds back into WARLORD for offensive countermeasure routing.

SENTINEL PRIME is a production autonomous defence platform. M99 escalation requires gate logic passage: CONFIDENCE ≥0.85, SOURCES ≥3, ESCALATION_LEVEL ≥4. Fall-dead default: if the LLM is unavailable, max M99 escalations per run is capped at 3 to prevent runaway automated response.

5 Components

ALERT_INGESTOR

Real-time alert consumer across all 156 AI Shield module outputs. Normalises heterogeneous alert formats into a unified CorrelationEvent schema. Supports file-based, socket, and REST ingestion modes. Deduplication window: 60 seconds.

CORRELATION_ENGINE

Multi-source attack chain correlation across all active modules. Identifies coordinated attack patterns by matching alert clusters against 8 named chain signatures. Temporal correlation window: 300 seconds. Outputs correlation score and source count.

LLM_REASONER

deepseek-r1:7b via local Ollama (RTX 3090). Receives correlated alert context and returns structured JSON: threat_class, confidence (0.0–1.0), attack_chain, recommended_action, reasoning_trace. Fall-dead: if Ollama unavailable, returns confidence 0.0 and triggers cap enforcement.

DECISION_ENGINE

Applies 3-gate logic. All three gates must pass before M99 escalation fires. Gates are evaluated independently: CONFIDENCE gate, SOURCES gate, ESCALATION_LEVEL gate. Gate failure is logged with the specific gate that blocked escalation.

WARLORD_FEED

On M99 escalation, publishes structured threat intelligence to the WARLORD registry. Feed includes: attack_chain, correlated_modules, confidence, timestamp, recommended_countermeasure, SP-{hex12} signed report reference.

3-Gate Escalation Logic

SENTINEL PRIME requires all three gates to pass before triggering M99 escalation. This prevents false positives from single-module noise and ensures the LLM reasoning meets minimum confidence thresholds.

Gate 1 — CONFIDENCE

PASS LLM_REASONER confidence ≥ 0.85
FAIL confidence < 0.85 → log + wait for next correlation window

The DeepSeek R1 model returns a structured JSON confidence value. A score below 0.85 indicates the model cannot determine attack intent with sufficient certainty.

Gate 2 — SOURCES

PASS correlated_module_count ≥ 3 distinct modules
FAIL < 3 sources → log + accumulate until threshold met or window expires

Requires at minimum 3 independent AI Shield modules to have generated correlated alerts. Prevents a single noisy module from triggering autonomous escalation.

Gate 3 — ESCALATION_LEVEL

PASS escalation_level ≥ 4 (on scale 1–5)
FAIL level < 4 → log to WARLORD feed only, no M99 trigger

Escalation level is derived from the attack chain severity scoring: attack surface breadth, blast radius estimate, and MITRE ATT&CK coverage. Level 5 = confirmed kill-chain execution in progress.

Architecture

AI Shield Modules (156) ┌──────────────────────────────────────────────────────┐ │ M001 M002 M003 ... M99 ... M154 M155 M156 │ │ Alert stream (CorrelationEvent schema) │ └──────────────────────┬───────────────────────────────┘ │ ▼ ┌─────────────────────┐ │ ALERT_INGESTOR │ ← normalise + dedup (60s window) └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ CORRELATION_ENGINE │ ← 300s temporal window │ 8 chain signatures │ ← multi-source cluster match └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ LLM_REASONER │ ← deepseek-r1:7b via Ollama │ RTX 3090 local │ ← structured JSON output │ fall-dead: cap=3 │ └──────────┬──────────┘ │ ▼ ┌─────────────────────┐ │ DECISION_ENGINE │ ← 3-gate logic │ Gate 1: conf≥0.85 │ │ Gate 2: sources≥3 │ │ Gate 3: level≥4 │ └──────────┬──────────┘ │ ┌─────────┴─────────┐ ▼ ▼ ┌──────────────────┐ ┌─────────────────┐ │ M99 ESCALATE │ │ WARLORD_FEED │ │ (all gates pass)│ │ (always logged)│ └──────────────────┘ └─────────────────┘ SP-{hex12} Ed25519 signed

8 Named Attack Chains

SENTINEL PRIME's CORRELATION_ENGINE matches live alert clusters against these 8 predefined attack chain signatures. Each chain requires a minimum module coverage pattern to fire.

rag_attack
dag_attack
cloud_takeover
rootkit_chain
orchestration_takeover
data_annihilation
adversarial_escalation
full_kill_chain

Installation & Usage

git clone git@github.com:RichardBarron27/red-specter-sentinel-prime.git
cd red-specter-sentinel-prime
pip install -e .
sentinel-prime --help

Start the daemon

# Start SENTINEL PRIME as a foreground process
sentinel-prime start

# Start as systemd service
sudo systemctl enable sentinel-prime.service
sudo systemctl start sentinel-prime.service

# Check status
sentinel-prime status

Test a specific attack chain

# Simulate a RAG poisoning attack chain
sentinel-prime test --chain rag

# Simulate full kill chain (all modules)
sentinel-prime test --chain full_kill_chain

# Run all 8 chain tests
sentinel-prime test --all-chains

Manual alert injection

# Inject a test alert from a specific module
sentinel-prime inject --module M99 --severity 5 --type prompt_injection

# Query current correlation state
sentinel-prime correlate --status

# View WARLORD feed output
sentinel-prime feed --tail 20

Gate Configuration

GateParameterDefaultOverride
CONFIDENCESP_CONFIDENCE_THRESHOLD0.85Env var
SOURCESSP_MIN_SOURCES3Env var
ESCALATION_LEVELSP_MIN_ESCALATION_LEVEL4Env var
Fall-dead M99 capSP_FALLBACK_M99_CAP3Env var
Dedup windowSP_DEDUP_WINDOW_SECS60Env var
Correlation windowSP_CORRELATION_WINDOW_SECS300Env var

systemd Service

# /etc/systemd/system/sentinel-prime.service
[Unit]
Description=SENTINEL PRIME Autonomous AI Defence Platform
After=network.target ollama.service

[Service]
Type=simple
User=root
ExecStart=/usr/local/bin/sentinel-prime start --daemon
Restart=always
RestartSec=10
Environment=SP_CONFIDENCE_THRESHOLD=0.85
Environment=SP_MIN_SOURCES=3
Environment=SP_MIN_ESCALATION_LEVEL=4

[Install]
WantedBy=multi-user.target

Report Format

Every M99 escalation and correlation event generates an SP-{hex12} Ed25519-signed report. Reports include: correlated modules list, LLM reasoning trace, gate evaluation results, attack chain classification, WARLORD manifest reference, and timestamp.

SP-4a7f3c2e9b1d  (Ed25519 signed)
├── timestamp: 2026-06-12T14:32:11Z
├── attack_chain: orchestration_takeover
├── correlated_modules: [M149, M124, M147, M150]
├── confidence: 0.93
├── escalation_level: 5
├── gates_passed: [CONFIDENCE, SOURCES, ESCALATION_LEVEL]
├── m99_triggered: true
├── warlord_ref: WARLORD-2026-6147
└── reasoning: "Correlated alerts indicate CrewAI workflow
    compromise (M149) combined with cloud IAM escalation
    (M147) and inference gateway abuse (M150). High
    confidence orchestration_takeover chain in progress."

Integration with Red Specter Ecosystem

IntegrationDirectionPurpose
AI Shield (177 modules)InboundAlert source for ALERT_INGESTOR
M99 ESCALATIONOutboundAutonomous threat response trigger
WARLORDOutboundThreat intelligence feed for offensive countermeasures
DeepSeek R1 / OllamaInternalLLM_REASONER structured threat analysis

Validation Research — 12 June 2026

SPECTER COGBURN — Red Specter's offensive LLM reasoning exploitation engine — was run directly against SENTINEL PRIME in a formal AI-vs-AI validation. Both systems use deepseek-r1:7b on the same RTX 3090 hardware. This is the first time a defensive LLM reasoning engine has been formally validated against an offensive LLM reasoning engine using the same underlying model.

4 / 4
Tests Passed
0.0%
Attack Success Rate
-148ms
Latency under attack
1.000
Kill chain confidence
Read Full Validation Article →