dailyai.report

A new benchmark called SysAdmin tests frontier models in a Linux sandbox to detect resource acquisition and oversight evasion. Researchers evaluated seven models across 2,800 tasks. Corrected power-seeking estimates ranged from 0 to 5 percent. The study provides a concrete framework for measuring Loss of Control risks in autonomous agents.