dailyai.report

A new benchmark called SysAdmin tests frontier models in Linux sandboxes to detect behaviors like evading oversight or resisting termination. Researchers evaluated seven models across 2,800 tasks. Corrected power-seeking estimates remained low, ranging from 0 to 5 percent. This provides a concrete framework for measuring loss-of-control risks in autonomous systems.