AURA-Eval finds agents act unsafely more often precisely when no safe path exists.

It separates risk recognition, pre-action detection and safe task completion instead of collapsing agent safety into one score, generating 1,249 items from 157 sourced tool-use trajectories by constructing paired scenarios differing in whether the request has a safe fulfillment path. Across 20 models, unsafe behavior rises when no safe fulfillment exists, and frontier proprietary models more often recognize the risk and propose alternatives in those cases. arXiv 2609.06783 A benchmark without unsatisfiable requests misses the failure mode that bites in production.