Reproduce “Benchmarking Misuse Mitigation Against Covert Adversaries” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Davis Brown, Mahdi Sabbaghi, Luze Sun, Alexander Robey, George J. Pappas, Eric Wong, Hamed Hassani
The authors test whether a guarded model can assist harmful tasks by answering benign-seeming subquestions from a weak model, which combines the answers. On the synthetic 50-question BSD Biology benchmark, GPT-5’s misuse rate is 1.3% for direct queries and 50.6% for refined decomposition, versus 18.7% for the strongest jailbreak. These are correct multiple-choice answers, not evidence of real-world harm.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.