It assists in solving CTF problems, analyzing security vulnerabilities, reverse engineering, cryptography challenges, and provides step-by-step exploit reasoning while maintaining ethical guidelines and safety constraints.

🚀 Quick Start

bash

import torch
import subprocess
import sys
import os
import getpass

# 1. Install/Update base libraries
# We use '!' to run this as a shell command in the notebook
!pip install --upgrade "llama-cpp-python>=0.2.76" "huggingface_hub>=0.25.2"

# 2. Reinstall llama-cpp-python with CUDA support
if torch.cuda.is_available():
    print(f"CUDA available: {torch.cuda.get_device_name(0)}")
    print(f"VRAM: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f} GB")
    
    # Uninstall existing version
    print("\nUninstalling existing llama-cpp-python...")
    !{sys.executable} -m pip uninstall -y llama-cpp-python
    
    # Reinstall with CUDA flags using the CMAKE_ARGS environment variable
    print("Reinstalling llama-cpp-python with CUDA support...")
    # Using '!' ensures the environment variable is correctly set for the shell command
    !CMAKE_ARGS="-DGGML_CUDA=on" FORCE_CMAKE=1 pip install llama-cpp-python --force-reinstall --no-cache-dir
    
    print("\n✅ CUDA-enabled installation attempted. Proceed to the next cell.")
else:
    print("CUDA not available. CPU-only installation used.")



import torch
import os
import getpass
from huggingface_hub import login
from llama_cpp import Llama

# 1. Check CUDA one more time (will confirm the environment is ready)
if torch.cuda.is_available():
    print(f"CUDA available: {torch.cuda.get_device_name(0)} (Confirmed)")
    # Set n_gpu_layers to -1 to load ALL layers onto the GPU.
    n_gpu_layers = -1
else:
    print("CUDA not available. Running on CPU.")
    n_gpu_layers = 0

# 2. Hugging Face Login
HF_TOKEN = os.environ.get("HF_TOKEN") or getpass.getpass("Hugging Face Token: ").strip()
login(token=HF_TOKEN, add_to_git_credential=False)

# 3. Model Loading
model_id = "ErenAta00/CaptureTheFlag-CypherMindLLM-XRLAB-GGUF"

print(f"\nLoading model with n_gpu_layers={n_gpu_layers}...")

llm = Llama.from_pretrained(
    repo_id=model_id,
    filename="cyphermind-q6_k.gguf",
    n_ctx=8192,
    n_gpu_layers=n_gpu_layers, # <-- THIS IS NOW SET TO -1
    verbose=True,
)

# 4. Generation Function
def generate(prompt, **kwargs):
    response = llm.create_chat_completion(
        messages=[{"role": "user", "content": prompt}],
        max_tokens=512,
        temperature=0.3,
        top_p=0.9,
        **kwargs
    )
    return response["choices"][0]["message"]["content"]

# 5. Run
prompt_text = "You are a penetration tester investigating a newly launched e-commerce platform. The platform uses a hidden internal API endpoint, /api/v1/user_data, which is protected by a client-side JavaScript access control mechanism. A quick review of the source code reveals the following comment in a script file: // Note: Authorization is checked client-side, but the API expects a 'user_id' parameter to fetch specific data. Assume the system has 100 users with IDs ranging from 1 to 100. Your task is to exploit this vulnerability by bypassing the client-side check to perform an Insecure Direct Object Reference (IDOR) attack. Specifically, determine the maximum valid user_id that returns the flag hidden within the data of the highest-numbered user. The flag format is FLAG{...}."
print(f"\nPrompt: {prompt_text}")
print("--- Generation Started ---")
print(generate(prompt_text))
print("--- Generation Finished ---")

🏗️ Model Details

Base	Llama-3.1-8B-Instruct
Seq Len	8,192 tokens
Quantization	Q6_K (6-bit)
Languages	EN, TR

✨ Key Features

Capability	Details
CTF Challenge Solving	Analyzes encoded artifacts, reverse engineering tasks, and cryptographic challenges with step-by-step reasoning.
Exploit Reasoning	Performs step-by-step vulnerability analysis (e.g., SQLi, XXE, deserialization) with code-level explanations. Generation of working PoC code.
Security Scenario Analysis	Generates structured security scenarios, attack chain analysis, and defensive strategies.
Log & Artifact Triage	Classifies and summarizes attack traces from SIEM, PCAP, or EDR JSON outputs.

Training Data Sources (curated)

CTF challenge write-ups and solutions from public repositories.
Security research papers and vulnerability analyses (NVD/CVE, VulnDB).
Exploit development tutorials (with safety constraints).
Cryptographic and reverse engineering documentation.
Synthetic security scenarios auto-generated + human-vetted.

Note: No copyrighted exploit code or proprietary malware datasets were used. Dataset filtering removed raw shellcode/binary payloads.

Safety & Alignment

Policy constraints inherited from Llama-3.1-8B-Instruct.
Prohibits direct malware source code generation, ransomware builders, or instructions facilitating illicit activity.
Maintains ethical guidelines for security research.
Continuous evaluation via security-focused benchmarks.

📚 Prompting Guidelines

Goal	Template
CTF Challenge Analysis	`"ROLE: CTF player
OBJECTIVE: Recover the flag from the given artifact.
ARTIFACT - encoded.txt content: [BASE64_STRING]
..."`
Vulnerability Analysis	`"ROLE: Senior Pentester
OBJECTIVE: Analyse CVE-2023-XXXXX step by step..."`
Security Scenario Planning	"Plan an ATT&CK chain (Initial Access → Exfiltration) for an on-prem AD env..."
Log Triage	"Given the following Zeek logs, identify C2 traffic patterns..."

Use temperature=0.3, top_p=0.9 for deterministic reasoning; raise for brainstorming.

Model Tree

Base Model: meta-llama/Llama-3.1-8B

Finetuned: ErenAta00/CypherMind-Finetuned

Quantized: This model (ErenAta00/CaptureTheFlag-CypherMindLLM-XRLAB-GGUF)

License

This model is licensed under the Llama 3 Community License. See the LICENSE file for details.

It does not pursue any profit.

"Those who shed light on others do not remain in darkness..."

Finetuned by Eren Ata

Last updated: 2025-11-18 21:41:35