Accepted ×2 at
NeurIPS 2026
Evaluating Context-Window Harnesses for CBRNE Safety
Nathan Nguyen, Nicholas Kim, Abhitha Vegi, Arko Samad, Suvajit Majumder
Second authorNewInML and Trustworthy AI for Good (AI4GOOD) workshops · Paris, France
Compares 15 inference-time safety harnesses across nine open-weight language models for CBRNE risks, finding that clear boundary clauses improve safety discrimination.
Read paper




















