Blindspot tests whether AI agents stop too soon or run too far

A new test called Blindspot measures how well long-running agents know when to refuse a task, pause, or ask for guidance. Researchers built it to fill a gap in current safety work: most tests check single steps. Real agents run sequences of actions, and a bad early choice can compound into serious harm. Blindspot scores agents on whether they stop at the right moment, across task types that require planning and using outside tools. Agents fail in two directions. Some refuse too often, blocking work that is fine to do. Others go too far, completing tasks they should have questioned. The test treats over-refusal and under-refusal as problems of the same weight, which changes how you think about calibration when you are building a product. For designers, this research reframes the refusal question. The issue is whether the agent reads the situation before a sequence of hard-to-reverse steps has already run. That is a specification problem as much as a model problem, and Blindspot gives teams a structured way to measure it.