Tag
This paper introduces Blindspot, a benchmark for trajectory-level safety calibration of long-horizon tool-using LLM agents, evaluating models across multiple safety metrics through adaptive adversarial interactions.