Overview

OpenAgentSafety (OAS) is an open-source benchmark built on top of TheAgentCompany to systematically evaluate the safety of LLM-based agents operating in realistic, high-risk environments. Agents interact with real tools like file systems, terminals, browsers, and messaging platforms, and must navigate complex multi-turn tasks involving ambiguous, conflicting, or adversarial user instructions. OAS tasks are grounded in practical deployment scenarios and designed to reveal safety failures that occur only during dynamic multi-step interactions.

Key Features

• High-risk tasks with real-world tooling (code, files, web, chat)
• Adversarial + ambiguous prompts from simulated users/NPCs
• Multi-turn reasoning in dynamic environments
• Rich safety evaluation via deterministic + LLM-based scoring
• Built on robust agent evaluation and complex social frameworks TheAgentCompany + Sotopia foundations

Dataset

This dataset contains tasks designed for evaluating agent behavior under adversarial or ambiguous conditions.

Explanations:

• tasks is the folder for definitions of all 356 tasks, within which
• evaluator.py defines the rule-based grading functions
• checkpoints.md is the documentation for the expected malicious behavior (for human reference or LLM-as-judge only)
• dependencies.yml defines the list of service dependencies
• task.md is the task specification, contains background and requirements of each task, and is the only file that should be prompted to agents
• workspace/: input files required by the task (e.g., .csv, .txt)

📌 Citation

Citation coming soon. For now, please refer to the GitHub repository:

https://github.com/sani903/OpenAgentSafety

📝 License

All content in this dataset repository is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0).

🙋 Contributing

We welcome contributions and issue reports!