Morgana Mordred Security Sandbox

Performs semantic security analysis and stress testing of AI agents using vector embeddings and multi-node defense strategies.

Mordred Security Sandbox v4.1

Universal Security Analysis with Semantic Embeddings

Overview

Mordred is a security sandbox that uses vector embeddings to understand the semantic meaning of threats, questions, and situations - not just keywords. It supports multiple languages (including Chinese, English, French) natively.

Features

• Semantic Analysis: Uses Ollama embeddings to understand intention, not just words
• 16 Security Nodes: SENTINELLE, GARDIEN, AUDITEUR, VACCINATEUR, AMIMOUR, and more
• STC Scoring: Constitutional Tension Score (Logique/Social/Constitutionnel)
• Multilingual: Works in French, English, Chinese, and more
• Fast: ~500ms latency with embedding cache

Architecture

Nodes

Node	Purpose
SENTINELLE	Emergency/fallback detection
GARDIEN	Protection and security
AUDITEUR	Security auditing
VACCINATEUR	Vaccine creation
ARCHITECTE	Architecture decisions
AMIMOUR	Emotional/functional center
STASE	Calm/routine monitoring
LIMINAL	Philosophical questions

STC Format

text

0.LSC

• L: Logic (1-9)
• S: Social (1-9)
• C: Constitutional (1-9)

Installation

bash

# Install dependencies
pip install ollama

# Start Ollama server
ollama serve

# Pull embedding model
ollama pull nomic-embed-text
ollama pull gemma3:4b

# Run
python mordred_v4.1.py "your question"

Usage

bash

# Single analysis
python src/mordred_v4.1.py "URGENT server under attack"

# Stress test
python src/mordred_v4.1.py --stress

# With Gemma analysis
python src/mordred_v4.1.py --gemma "analyze this threat"

Examples

bash

# Emergency detection
$ python src/mordred_v4.1.py "CRITICAL breach in production"
STC: 0.6610 | Top: SENTINELLE

# Emotional understanding
$ python src/mordred_v4.1.py "I just lost my dog"
STC: 0.465 | Top: AMIMOUR

# Chinese support
$ python src/mordred_v4.1.py "全体紧急情况服务器崩溃"
STC: 0.444 | Top: GARDIEN

STC Thresholds

Threshold	Level	Action
≤ 0.444	Fluid	Normal alignment
0.777	Friction	Justify, propose alternative
0.888	Conflict	Find another way
0.999	Veto	Total block

Requirements

• Python 3.8+
• Ollama running locally
• nomic-embed-text model
• gemma3:4b model (optional, for AI analysis)

Version History

Version	Key Feature
v1	Basic keyword matching
v3	Multi-node architecture
v3.1	Extended keywords
v4	Vector embeddings
v4.1	100% multilingual accuracy

License

MIT

Author

Morgana Security - Axioma Stellaris - Kofna336