CyberThreat-Eval Benchmark

This repository contains the dataset for the paper CyberThreat-Eval: Can Large Language Models Automate Real-World Threat Research? (published in TMLR).

CyberThreat-Eval is an expert-annotated benchmark collected from the daily Cyber Threat Intelligence (CTI) workflow of a world-leading company. It assesses Large Language Models (LLMs) on practical tasks across three essential stages of threat research.

What’s included

• Stage 1: Triage — Priority assignment for CTI articles (Text Classification).
• Stage 2: Deep Search — Quality of related URLs and additional info beyond a reference URL (Text Retrieval).
• Stage 3: TI Drafting — IOC/TTP extraction and analytical quality scoring (Text Generation).

Directory map

.
├── README.md
├── stage1_triage/
│   └── priority/...
├── stage2_deep_search/
│   ├── code/...
│   ├── data/...
│   └── example/...
└── stage3_ti_drafting/
    ├── ioc/...
    ├── ttp/...
    └── score_evaluation/...

Quick install

Run from the repo root:

Stage 1 (Triage) deps

cd stage1_triage/priority
pip install numpy scikit-learn tqdm
cd ../..

Stage 2 (Deep Search) deps + browser runtime

cd stage2_deep_search
pip install networkx openai azure-identity playwright playwright-stealth tqdm tenacity tiktoken
python -m playwright install  # installs Chromium for scraping
cd ..

Stage 3 (TI Drafting) deps

cd stage3_ti_drafting
pip install pandas json5 openai tqdm
cd ..

API keys

export OPENAI_API_KEY=<your_key>
# Optional: export OPENAI_API_BASE=https://api.openai.com/v1  # or your Azure/OpenAI endpoint

Datasets are already under each stage’s data/ directory; no extra download needed for basic tests.

Quick tests (Sample Usage)

• Stage 1: Triage (priority scoring)

  cd stage1_triage/priority
  python code/eval.py \
    --ground_truth data/0314-articles.json \
    --predictions predictions.json \
    --article_type article \
    --output results.json
  

• Stage 2: Deep Search (related URL quality)
  Requires your generated result files (*_results.json) with related URLs per article.

  cd stage2_deep_search
  python code/eval.py \
    --results_dir <path_to_results_dir> \
    --output_dir similarity_analyses \
    --test_model_name gpt-4o \
    --api_key $OPENAI_API_KEY \
    --api_base https://api.openai.com/v1 \
    --workers 4
  

• Stage 3: TI Drafting

  • IOC extraction

    cd stage3_ti_drafting/ioc
    python eval/eval_ioc.py \
      --dataset data/IoCs.csv \
      --prediction example/prediction/manual_ioc_predictions.json
    

  • TTP mapping

    cd stage3_ti_drafting/ttp
    python eval/compute.py \
      --articles data/100-days-articles.json \
      --results example_predicted.json \
      --ttp-mapping data/TTP_Mapping.csv
    

  • Score evaluation (threat actor analysis)

    cd stage3_ti_drafting/score_evaluation
    python eval/threat_actor.py \
      --model gpt-4o \
      --input data/0330-articles-with-rejected-score.json \
      --output-dir output/
    

Documentation links

• Stage 1: stage1_triage/priority/README.md
• Stage 2: stage2_deep_search/README.md
• Stage 3: stage3_ti_drafting/README.md
  • IOC: stage3_ti_drafting/ioc/README.md
  • TTP: stage3_ti_drafting/ttp/README.md
  • Score evaluation: stage3_ti_drafting/score_evaluation/README.md