Scientific Safety Levels (SSL)

A Governance Framework for Agentic Science in the Age of Autonomous Discovery

Read the PDF

Summary

This paper proposes Scientific Safety Levels (SSL), a governance framework for Agentic Science systems: AI systems that can move beyond summarization into hypothesis generation, experimental design, tool use, laboratory interaction, and iterative scientific workflows.

• SSL-1: minimal-risk scientific AI with read-only or tightly bounded use
• SSL-2: proposal-capable assistant with human-gated execution
• SSL-3: controlled scientific agent with bounded autonomous execution
• SSL-4: high-risk autonomous scientist in dual-use or catastrophic-risk settings
• SSL-5: systems requiring maximum containment or moratorium due to verifiability or capability-overhang concerns

The central claim is that existing frameworks such as BSL, ASL, and traditional research ethics do not fully govern scientific AI once it is deployed with tools, permissions, infrastructure access, and real-world actuation.

Key Contributions

• A five-level deployment framework for governing autonomous scientific systems.
• Formal trigger conditions based on a six-tuple deployment-state representation.
• A risk taxonomy spanning chem-bio, epistemic, autonomy, and cyber-physical risks.
• Technical control architectures, case studies, and certification pathways for audit and governance.

Release Note

This repository is intentionally minimal and contains the paper PDF together with a short release README.

Citation

If you use or reference this work, please cite it as:

@misc{xu_scientific_safety_levels,
  title = {Scientific Safety Levels (SSL): A Governance Framework for Agentic Science in the Age of Autonomous Discovery},
  author = {Xingcheng Xu and Chang Jin and An Wang and Qiaosheng Zhang and Chao Yang and Jingjing Qu and Xin Li and Wenwei Zhang and Kai Chen and Lei Bai and Xiangnan He and Yi Chang and Shiliang Zhang and Xiuquan Li and Nenghai Yu and Xianglong Liu and Kui Ren and Tiejun Huang and Liang He and Xia Hu and Bowen Zhou},
  institution = {Shanghai Artificial Intelligence Laboratory and University of Science and Technology of China and Jilin University and Peking University and Beihang University and Zhejiang University and Chinese Academy of Science and Technology for Development},
  url = {https://github.com/AI45Lab/ScientificSafetyLevels}
}