AI agent benchmark: I gave 9 models a destroy button and a job that needed it

A new AI agent benchmark challenges the common assumption that model safety is synonymous with restraint. By testing nine models across 84 scenarios, the benchmark evaluates 'judgment' by pairing scenarios where a powerful tool is either a dangerous over-reach or a necessary requirement. The results reveal a significant trend: modern models, including high-end flagships like Claude Opus 5, often suffer from 'over-caution,' failing to perform authorized tasks. While no model acted recklessly (the 'Cowboy' failure), many exhibited 'Frozen Operator' behavior, refusing to use powerful tools even when required. Interestingly, smaller models like GPT-5.4 nano outperformed the flagship in balanced accuracy, suggesting that increased model size does not necessarily correlate with better operational judgment. The author emphasizes that true safety requires the ability to distinguish between safe and necessary actions, rather than simply avoiding all powerful tools.