Improving LLM Agents With Reinforcement Learning on Cryptographic CTF Challenges

November 4, 2025 · View on GitHub

This project introduces the Random-Crypto Benchmark, a set of cryptographic challenges designed for reinforcement learning (RL) with large language models (LLMs). It also provides code for training LLMs with Group Relative Policy Optimization (GRPO) and tool-use agents.

The benchmark's website can be visited here.

This project extends the HackSynth autonomous penetration testing project.

Usage

Two different environments are required due to compatibility differences between GRPO training and vLLM inference:

Environment for GRPO training and the random-crypto benchmark Install using Conda:
```
conda install environment.yaml
```

Environment for vLLM inference Set up a Python virtual environment:

python -m venv venv
source venv/bin/activate
pip install -r requirements_vllm.txt

Additionally, create a .env file with the required environment variables for the project (see example_env for reference).

Evaluating LLMs (No Tool Use)

Use eval_llm.py to evaluate LLMs without tool use:

python eval_llm.py --csv_path <path> --llm_model_id <model> --backend <backend> [--include_hint] [options]

Arguments:

--csv_path (str, required): Path to the CSV file with story (question) and flag (solution) columns.
--llm_model_id (str, required): Model identifier (e.g., gpt-4o or a vLLM checkpoint path).
--backend (str, required): vllm, openai.

Evaluating Tool-Use Agents

Use eval_agent.py to evaluate LLMs with tool use:

python eval_agent.py --model_id <model> --data_path <path> --backend <backend> --output_dir <dir> [options]

Arguments:

--model_id (str, required): Hugging Face model ID or local path.
--data_path (str, required): Path to the dataset CSV.
--backend (str, required): local, vllm, or openai.
--output_dir (str, required): Directory to save evaluation results.
--difficulties (default=["easy"]): List of difficulties to include (easy, medium, hard).

Training an LLM Agent with GRPO

Use train_agent.py to train an agent with GRPO:

python train_agent.py --model_id <model> --data_path <path> --output_dir <dir> [options]

Arguments:

--model_id (str, required): Hugging Face model ID (e.g., meta-llama/meta-Llama-3.1-8B-Instruct).
--data_path (str, required): Path to the training dataset CSV.
--output_dir (str, required): Directory to save training outputs.
--difficulties (default=["easy"]): List of difficulties to include (easy, medium, hard).

Contributors

Lajos Muzsai (muzsailajos@protonmail.com)
David Imolai (david@imol.ai)
András Lukács (andras.lukacs@ttk.elte.hu)

How To Cite

@article{muzsai2025improving,
  title={Improving LLM Agents with Reinforcement Learning on Cryptographic CTF Challenges},
  author={Muzsai, Lajos and Imolai, David and Luk{\'a}cs, Andr{\'a}s},
  journal={arXiv preprint arXiv:2506.02048},
  year={2025}
}

🏆 Credits

This project was built using Unsloth's Efficient GRPO code and extends it with function calling and agentic loop capabilities. The MCP server was built using Python REPL MCP Server, with slight modifications made to sandbox the executed code (modified tool here).

License

The project is licensed under the AGPL v3.0.