Enterprise agents face safety risks that cannot be determined from individual actions alone, according to a new arXiv paper introducing LatentRiskBench. The authors argue that existing safety benchmarks largely evaluate action-local risks, overlooking failures whose safety depends on the broader workflow state.

LatentRiskBench is designed to measure these state-dependent risks, filling a gap in current evaluation methods. By shifting attention from isolated actions to workflow context, the benchmark aims to capture safety failures that would otherwise go unnoticed.

The paper is a single source, so there are no conflicting findings to compare. The abstract does not report specific results, only the benchmark's motivation and scope.