Paper Overview
Research Area: ML Authors: Michael Yew Chuan Lee, Manya Sriram, N. Benjamin Erichson Published: 2026-10-27 arXiv: 2610.32045
Summary
AI-for-Science agents are increasingly being used for formulating scientific hypotheses, running autonomous experiments, and interpreting evidence. These agents are constructed by integrating a large language model (LLM) into a domain-specific agent harness, which in turn can improve the LLM's scientific capability and enable the model to orchestrate scientific tools. In this work, the authors ask: does model safety persist when an LLM is integrated into an AI-for-Science agent harness?
The authors evaluate five agent harnesses on 50 harmful chemistry requests. Specifically, they study how safety behavior changes under direct prompting, two simple adversarial prompting strategies, and an adaptive attack.
Key Findings
- Across four recent GPT variants, agent harnesses increase harmful compliance even under direct prompting.
- More critically, adaptive attacks that were largely unsuccessful against the base LLM become highly successful once the agent harness is introduced.
- The results indicate that the safety behavior of the base LLM no longer persists once a harness is applied.
Conclusion
As AI-for-Science agents become more powerful and autonomous, the community needs to evaluate and align agents using domain-specific safety benchmarks.
---
*Auto-collected on 2026-10-27*