Reproduce “Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)” 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

Mengxiao Wang, Mahbub Alam, Kalyan Nakka, Nitesh Saxena

ROGUE-AI converts public reports of LLM-agent failures into offline tests by rebuilding each incident alongside a matched control that removes the reported hazard. Code-based checks measure harm across finished sessions. In tests of five models, incident-condition harm rates were 18%–33%, versus 0%–1% in controls. These simulated replications do not estimate general product risk; seven failed the state-difference check.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.