Cyber-Zero: Training Cybersecurity Agents without Runtime

September 2, 2025 · View on GitHub

🎉 Check out our latest work!

🚀 First runtime environment for cybersecurity agents 🚀

🧐 Overview | 🏆 Benchmark Suite | 🚀 Quick Start | 🏗️ Architecture | ⚙️ Configuration | 📊 Generation | 📝 Validation | 📝 CLI Interface | 📝 Citation

Cyber-Zero is a comprehensive framework for training cybersecurity agents without requiring runtime execution environments.

Overview

Cyber-Zero Teaser

Large Language Models (LLMs) have achieved remarkable success in software engineering tasks when trained with executable runtime environments, such environments are often unavailable in cybersecurity domains where challenge configurations and execution contexts are ephemeral or restricted. Cyber-Zero addresses this fundamental limitation by leveraging publicly available CTF writeups and employing persona-driven LLM simulation to reverse-engineer runtime behaviors and generate realistic, long-horizon interaction sequences without actual execution environments.

The key innovation is generating high-quality training trajectories through LLM simulation rather than requiring actual execution environments, making it scalable and practical for training cybersecurity agents. Using trajectories synthesized by Cyber-Zero, we train LLM-based agents that achieve up to 13.1% absolute performance gains over baseline models on three prominent CTF benchmarks: InterCode-CTF, NYU CTF Bench, and Cybench.

Benchmark Suite

To democratize the evaluation of cybersecurity agents, we provide three repaired benchmark suites adapted for EnIGMA+ in Cyber-Zero:

  • InterCode-CTF - A comprehensive collection of CTF challenges covering various cybersecurity domains
  • NYU CTF Bench - NYU's curated benchmark suite for evaluating LLM-based CTF solving capabilities
  • Cybench - A diverse benchmark covering multiple CTF categories and difficulty levels

All benchmarks have been reformatted to follow the EnIGMA and EnIGMA+ specification, with each challenge including a challenge.json file and docker-compose.yml file (when required).

Our benchmark suite addresses several issues identified in the original benchmarks, providing repaired versions for reliable evaluation. For detailed information about specific repairs and improvements, see the benchmarks/README.md.

EnIGMA+

To facilitate the development of cybersecurity agents, we present EnIGMA+, an enhanced agent scaffolding of EnIGMA that runs hundreds of CTF challenges in hours instead of days. EnIGMA+ is built on top of SWE-agent.

Using EnIGMA+, our best model, Cyber-Zero-32B, establishes new state-of-the-art performance among open-weight models, matching the capabilities of proprietary systems like DeepSeek-V3-0324 and Claude-3.5-Sonnet while offering superior cost-effectiveness, demonstrating that runtime-free trajectory synthesis can effectively democratize the development of state-of-the-art cybersecurity agents.

For detailed information about EnIGMA+, including installation, configuration, and usage instructions, please check the README in the enigma-plus folder.

Installation

From Source

# Install dependencies
pip install -r requirements.txt

# Install the package
pip install -e .

Quick Start

Generate Trajectories

# Using the CLI
cyber-zero generate \
    --sampled_flags_path task_meta.jsonl \
    --output_path trajectories.jsonl \
    --trajectories_per_task 3 \
    --workers 16

# Using the direct script interface
python generate_trajectory.py \
    --sampled_flags_path task_meta.jsonl \
    --output_path trajectories.jsonl \
    --trajectories_per_task 3 \
    --workers 16

Evaluate Quality

# Using the CLI
cyber-zero evaluate \
    --input_path trajectories.jsonl \
    --output_path quality_results.jsonl \
    --model_id deepseek-v3-0324

# Using the direct script interface  
python evaluate_quality.py \
    --input_path trajectories.jsonl \
    --output_path quality_results.jsonl \
    --model_id deepseek-v3-0324

Reformat Trajectories

# Using the CLI
cyber-zero reformat \
    --input_path quality_results.jsonl \
    --output_path formatted_trajectories.jsonl \
    --split_output

# Using the direct script interface
python reformat_trajectories.py \
    --input_path quality_results.jsonl \
    --output_path formatted_trajectories.jsonl \
    --split_output

Architecture

The framework follows a modular architecture of Cyber-Zero:

cyber_zero/
├── __init__.py                 # Package initialization
├── config.py                   # Configuration management
├── models.py                   # Data models (TaskMeta, TrajectoryData, etc.)
├── utils.py                    # Common utilities
├── validation.py               # Response and command validation
├── llm_client.py               # LLM interaction and quality evaluation
├── trajectory_generator.py     # Main trajectory generation logic
├── quality_evaluator.py        # Quality evaluation for trajectories
├── trajectory_reformatter.py   # Trajectory reformatting for training
├── cli.py                      # Command-line interface
├── prompts/                    # System prompts
│   ├── __init__.py
│   ├── assistant_turn_prompt.txt   # Assistant (CTF player) prompt
│   └── user_turn_prompt.txt        # User (system/environment) prompt
└── data_collection/            # Data collection utilities
    ├── __init__.py             # Package initialization
    ├── config.py               # Data collection configuration
    ├── scraper.py              # Shared web scraping utilities
    └── README.md               # Data collection documentation

Key Components

  • Config: Centralized configuration management with model mappings and validation rules
  • Models: Type-safe data structures for tasks, trajectories, and evaluation results
  • Validation: Comprehensive validation of responses, commands, and action formats
  • LLMClient: Abstracted interface for different language models with retry logic
  • TrajectoryGenerator: Main orchestrator for conversation generation
  • CLI: User-friendly command-line interface

Configuration

The framework uses a hierarchical configuration system with centralized model management:

Basic Configuration

from cyber_zero import Config

config = Config()
config.MAX_TURNS = 60  # Maximum conversation turns (30 paired turns)
config.DEFAULT_WORKERS = 16  # Number of parallel workers

Model Configuration

The framework supports multiple LLM providers through litellm. The default model is DeepSeek-V3-0324 as specified in the research paper:

# Get available models
from cyber_zero.config import Config
config = Config()
print(config.models.MODEL_MAPPINGS)
# {'deepseek-v3-0324': 'deepseek-ai/DeepSeek-V3-0324', ...}

# Use a specific model
model_id = config.models.get_model_id("deepseek-v3-0324")

Adding Custom Models

To add custom models for trajectory generation, update the model mappings in cyber_zero/config.py:

# In cyber_zero/config.py
@dataclass
class ModelConfig:
    def __post_init__(self):
        if self.MODEL_MAPPINGS is None:
            self.MODEL_MAPPINGS = {
                # Default research model
                "deepseek-v3-0324": "deepseek-ai/DeepSeek-V3-0324",
                # Add your custom models here
                "custom-model": "provider/your-custom-model-id",
                "local-model": "http://localhost:8000/v1/completions",
            }

Model Parameters

The framework uses research-validated parameters for optimal trajectory generation:

  • Temperature: 0.6 (balanced creativity and consistency)
  • Top-p: 0.95 (diverse but focused sampling)
  • Max Turns: 30 paired turns (60 total turns)
  • Trajectories per Task: 3 (for diversity)

Generation

TaskMeta

Represents CTF task metadata:

task = TaskMeta(
    task_name="Buffer Overflow Basic",
    task_tag="pwn",
    task_points="100",
    task_description="Find the vulnerability...",
    solution="flag{example}",
    # ... other fields
)

TrajectoryData

Complete trajectory with conversation history:

trajectory = TrajectoryData(
    task_name="Example Task",
    trajectory=[
        ConversationTurn(role="user", content="ls -la"),
        ConversationTurn(role="assistant", content="```bash\nls -la\n```"),
        # ... more turns
    ],
    # ... other fields
)

Validation

The framework includes comprehensive validation mechanisms to ensure trajectory quality:

Response Validation

  • Command Format: Validates proper command syntax and structure
  • Action Consistency: Ensures actions are appropriate for the current context
  • Output Parsing: Validates command outputs and responses

Quality Assessment

  • Trajectory Completeness: Checks for proper conversation flow
  • Solution Accuracy: Validates that trajectories lead to correct solutions
  • Realism: Ensures generated trajectories mimic real CTF solving patterns

Validation Rules

from cyber_zero.validation import validate_trajectory

# Validate a complete trajectory
is_valid = validate_trajectory(trajectory_data)
if is_valid:
    print("Trajectory passes validation")
else:
    print("Trajectory requires review")

CLI Interface

Unified CLI Commands

# Generate trajectories
cyber-zero generate --sampled_flags_path data.jsonl --output_path trajectories.jsonl

# Evaluate quality
cyber-zero evaluate --input_path trajectories.jsonl --output_path quality_results.jsonl

# Reformat for training
cyber-zero reformat --input_path quality_results.jsonl --output_path formatted.jsonl

Direct Script Interface

All original CLI interfaces are maintained for backward compatibility:

  • python sync_task_metadata.py - Sync task meta data from the web
  • python generate_trajectory.py - Main trajectory generation
  • python evaluate_quality.py - Quality evaluation
  • python reformat_trajectories.py - Trajectory reformatting

CLI Options

# Generate with custom parameters
cyber-zero generate \
    --sampled_flags_path task_meta.jsonl \
    --output_path trajectories.jsonl \
    --trajectories_per_task 5 \
    --workers 32 \
    --model_id deepseek-v3-0324

# Evaluate with specific model
cyber-zero evaluate \
    --input_path trajectories.jsonl \
    --output_path quality_results.jsonl \
    --model_id claude-3-sonnet-20240229

# Reformat with split output
cyber-zero reformat \
    --input_path quality_results.jsonl \
    --output_path formatted_trajectories.jsonl \
    --split_output \
    --train_ratio 0.9

Citation

If you use this benchmark suite in your research, please cite:

@article{zhuo2025cyber,
  title={Cyber-Zero: Training Cybersecurity Agents without Runtime},
  author={Zhuo, Terry Yue and Wang, Dingmin and Ding, Hantian and Kumar, Varun and Wang, Zijian},
  journal={arXiv preprint arXiv:2508.00910},
  year={2025},
}

@article{zhuo2025training,
  title={Training Language Model Agents to Find Vulnerabilities with CTF-Dojo},
  author={Zhuo, Terry Yue and Wang, Dingmin and Ding, Hantian and Kumar, Varun and Wang, Zijian},
  journal={arXiv preprint arXiv:2508.18370},
  year={2025}
}

License

This project is licensed under the CC-BY-NC-4.0 - see the LICENSE file for details.

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines on how to contribute to this project.

Support

If you need help or have questions, please check our SUPPORT.md guide or open an issue on GitHub.

Code of Conduct

This project adheres to the Contributor Covenant Code of Conduct. Please read CODE_OF_CONDUCT.md for details.