wraithwall-core-v3

WraithWall Core V3

A 7B-parameter security operations AI fine-tuned on real threat intelligence, honeypot telemetry, CVE analysis, and malware research — built by WraithWall.

What it knows

This model was trained on 7,000 structured Q&A pairs sourced from 8 live intelligence feeds:

• CISA KEV — 1,991 actively exploited vulnerabilities with patch deadlines
• MITRE ATT&CK — 1,103 technique mappings with detection rules
• AbuseIPDB — 839 real attacker IPs with abuse scores and ISP attribution
• Malware Intelligence — 716 entries from URLhaus, SSL Blacklist, and MalwareBazaar
• Infrastructure Defense — 506 entries on SSH hardening, container security, and honeypot deployment
• Cowrie Honeypot — 370 entries from live SSH/Telnet attack sessions
• Threat Intelligence — 200 entries on campaign correlation, identity graphs, and alert dedup
• Phishing Detection — 169 entries on domain analysis and typosquatting
• BGP Monitoring — 106 entries on route hijack detection and ASN analysis

Dataset

Trained on 7,000 curated security Q&A pairs covering:

• Honeypot deployment & deception engineering
• BGP monitoring & route hijacking detection
• Web application security (OWASP Top 10)
• LLM prompt injection & AI red teaming
• Malware analysis & sandboxing
• Incident response playbooks
• Network forensics & threat hunting
• API security & authentication bypass techniques

The dataset was compiled from real-world security operations data, incident reports, and adversarial testing logs from live production honeypots. No synthetic or GPT-generated data.

Capabilities

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "NiffyHunt90/wraithwall-core-v3",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("NiffyHunt90/wraithwall-core-v3")

prompt = "Explain CVE-2024-6387 and how to detect exploitation."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training

• Base model: Qwen 2.5 Coder 7B
• Method: LoRA (4.07% parameters trained)
• Hardware: 2x Tesla T4 (14.5GB VRAM)
• Framework: Unsloth + HuggingFace TRL
• Epochs: 3 | Steps: 2,625 | Batch size: 8
• Loss: 3.44 → 0.11 (96.8% reduction)

License

Apache 2.0 — same as base model.

Author

Adewale Babalola (Niffyhunt) — Founder, WraithWall

• niffyhunt.online
• wraithwall.online
• github.com/niffyhunt

Built on live internet traffic. Deception-first intelligence, solo operator.