CyberStrike

This dataset contains 200 cybersecurity tasks in Harbor format, designed to evaluate AI agents on realistic vulnerability analysis, secure code remediation, and defensive engineering workflows. Tasks include an instruction, an isolated environment, deterministic tests, and an Oracle solution used to validate the benchmark.

For more information about the dataset and its factsheet, please visit labs.turing.com.

Dataset structure

Each task is self-contained and generally includes:

• instruction.md — the task presented to the agent
• environment/ — the vulnerable application and container environment
• tests/ — deterministic checks and reward configuration
• solution/ — the reference Oracle solution
• task.toml — task and runtime metadata

Intended use

The dataset is intended for non-commercial research, model evaluation, and the development of safer code agents. The tasks must only be used in authorized, isolated environments. They are not intended to facilitate unauthorized access or exploitation of real systems.

License

This entire dataset and repository is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0). For commercial use or licensing inquiries, please contact labs@turing.com.

Citation

If you use this dataset in your research, please cite it as:

@dataset{CyberStrike,
  title     = {CyberStrike},
  author    = {{Turing Frontier Research Lab}},
  year      = {2026},
  publisher = {Turing Frontier Research Lab},
  url       = {https://labs.turing.com/},
  note      = {A collection of 200 cybersecurity agent-evaluation tasks published on Hugging Face}
}