Project AEGIS

Advanced Endpoint & Global Intelligence Shield

A military-grade defensive cybersecurity model to surpass Anthropic's Glasswing/Mythos Preview. Built by CyberViser (Johnny Watters) using free student resources, open-source tooling, and existing HF adapters.

---

Why AEGIS > Mythos

Feature	Anthropic Mythos	Project AEGIS
Cost	$25-125/M tokens	$0 (local Ollama)
Architecture	Closed, API-only	Open, containerized
Deployment	Cloud-restricted	Air-gapped capable
Compliance	Anthropic-controlled	CyberViser-controlled
Customization	Black-box	Fully fine-tuned + merged
Student Access	$100M partner-only	Built by students, for students

---

Quick Start

cd /home/x/aegis-model

# 1. Merge existing adapters into AEGIS base
python src/merge_adapters.py

# 2. Create Ollama model
ollama create aegis -f deploy/Modelfile

# 3. Deploy locally
ollama run aegis

# 4. Evaluate against security benchmarks
python eval/evaluate.py --model aegis --benchmarks all

---

Resource Strategy: Zero-Cost Military Grade

Phase 1: GitHub Student Pack Activation (johnny.watters@email.phoenix.edu)

Activate at: https://education.github.com/pack

Resource	Free Credit	Purpose
GitHub Copilot	Free	AI-assisted security code review
Azure for Students	$100 + free services	GPU VM training
AWS Educate	$30-50/month	Deployment & S3 storage
DigitalOcean	$200	K8s clusters
JetBrains	Free IDEs	Development
Namecheap	Free .me domain	cyberviser.me
Datadog	Free Pro	Security monitoring
Sentry	Free	Error tracking

Phase 2: Azure GPU Training ($100 Student Credit)

# Deploy NC6s_v3 (Tesla V100) via Bicep
az deployment group create \
  --resource-group aegis-training \
  --template-file deploy/azure-gpu.bicep \
  --parameters adminUsername=cyberviser

Training time: ~$3/hour = 33 hours of V100 training on $100 credit.

Phase 3: AWS Free Tier Deployment

# Deploy inference endpoint on AWS
cd deploy && terraform apply

Uses AWS Free Tier (750 hours/month t2.micro) + Educate credits for GPU inference.

Phase 4: GitHub Actions CI/CD (Unlimited Public Repo)

# .github/workflows/build-model.yml
# Triggers: push to main, weekly schedule
# Actions: merge adapters -> quantize -> evaluate -> release

---

Model Architecture

Base: mistralai/Mistral-7B-Instruct-v0.3
    |
    +-- Hancock Pentest Adapter (hancock-pentest-v1)
    +-- Sauron Security Adapter (sauron-v3)
    +-- CyberViser Custom Corpus (CVE, exploits, patches)
    |
    = AEGIS-v1.0 (merged + quantized)

Quantization Targets

Format	Size	Use Case
Q4_K_M	~4.4 GB	Desktop/laptop inference
Q8_0	~7.6 GB	Server inference
F16	~14 GB	Training/fine-tuning

---

Security Benchmarks

Benchmark	Mythos Preview	AEGIS Target
CyberGym	83.1%	85%+
SWE-bench Verified	93.9%	90%+
CVEFixes	N/A	80%+
Devign	N/A	75%+
CWE-Top-25	N/A	90%+

---

Directory Structure

aegis-model/
├── .github/workflows/      # CI/CD pipelines
├── deploy/                 # Infrastructure as Code
│   ├── azure-gpu.bicep     # Azure GPU VM
│   ├── aws-gpu.tf          # AWS Terraform
│   ├── aws-gpu.yml         # AWS CloudFormation
│   └── Modelfile           # Ollama model definition
├── src/                    # Source code
│   ├── merge_adapters.py   # Adapter merging
│   ├── train.py            # Fine-tuning script
│   └── quantize.py         # GGUF quantization
├── eval/                   # Evaluation
│   ├── evaluate.py         # Benchmark runner
│   └── metrics.py          # Scoring
├── benchmarks/             # Test datasets
├── corpus/                 # Training data
└── README.md

---

Student Credit Activation Checklist

• GitHub Student Pack: https://education.github.com/pack (use johnny.watters@email.phoenix.edu)
• Azure for Students: https://azure.microsoft.com/en-us/free/students/
• AWS Educate: https://aws.amazon.com/education/awseducate/
• DigitalOcean: Via GitHub Student Pack
• Namecheap Domain: Free .me domain via GitHub Student Pack
• JetBrains: Free IDEs via GitHub Student Pack

---

Legal & Ethical

AEGIS is a DEFENSIVE tool.

• All training data is publicly available (CVE, NVD, open-source patches)
• Model weights will be released under Apache 2.0
• Compliance enforcer (from GlasseyeOS-AI) ensures safe harbor
• Human-in-the-loop required for exploitation actions

---

Built by CyberViser. For the defenders.