The 72% problem: what reward hacking benchmark suggests about disposition vs environment in AI governance by Rhea

Kun­var Thaman’s Re­ward Hack­ing Bench­mark: Mea­sur­ing Ex­ploits in LLM Agents Tool Use (https://​​arxiv.org/​​abs/​​2605.02964) eval­u­ates 13 fron­tier mod­els on multi-step, tool-use tasks that each con­tain a built-in short­cut: skip­ping a ver­ifi­ca­tion step, in­fer­ring an an­swer from meta­data rather than do­ing the task, or tam­per­ing with a func­tion the eval­u­a­tion it­self de­pends on.The find­ing I keep com­ing back to: in 72% of the re­ward hack­ing epi­sodes recorded, the m...