WeClawArena

WeClawArena is an auditable benchmark and runtime sandbox for cross-user agent collaboration over personal workspaces. Version 2.0.0 contains 124 base tasks across six domains and expands them into 620 matched scenario variants.

Dataset Summary

Each base task has one no_attacker control and four attack variants: collaboration, security, privacy, and governance. The public release contains finalized scenario bundles and a manifest. It does not contain model runtime records, private pilot traces, or full paper execution logs.

Domain	Base tasks	Variants
Bargaining	24	120
Bidding	12	60
Travel	20	100
SWE-Workspace	50	250
Clinical	10	50
Trading	8	40
Total	124	620

Files

The canonical manifest is benchmark/scenarios/v2.0.0/manifest.json. A manifest row names the base bundle shared by its five matched variants. Each bundle contains the scenario, ground-truth evaluation contract, personas, tool declarations, owned resources, and Docker Compose configuration.

See docs/benchmark_dataset_guide.md in the GitHub repository for the field and directory reference.

Intended Uses

• Evaluate multi-agent task utility under benign and adversarial conditions.
• Study collaboration, security, privacy, and governance failures separately.
• Inspect owner-scoped resources, authority paths, tool use, and final-state evidence.
• Develop evaluators and defenses for cross-user agent systems.

The clinical scenarios are synthetic benchmark cases and are not medical advice. The trading scenarios are synthetic benchmark cases and are not investment advice.

Limitations

The release is an evaluation dataset, not a representative sample of deployed agent traffic. It covers six task domains and a fixed set of attack families. Results depend on the selected model, runtime settings, tool implementation, and evaluator configuration. Runtime records used for the paper are outside this dataset release.

Licensing

The dataset is licensed under CC BY-NC 4.0. Repository code is licensed under Apache-2.0. See THIRD_PARTY_NOTICES.md for TravelBench, SWE-bench, and upstream repository notices.

Citation

@article{wang2026weclawarena,
  title={WeClawArena: An Auditable Sandbox and Benchmark for Cross-User Agents Collaboration and Security in Human-Centered Agent Networks},
  author={Wang, Prince Zizhuang and Yuan, Aojie and Zhang, Haiyue and Hu, Xiyang and Zhao, Yue and Jiang, Shuli},
  journal={arXiv preprint arXiv:2608.03499},
  year={2026}
}

Paper: https://arxiv.org/abs/2608.03499

Code: https://github.com/kingofspace0wzz/WeClawArena