dailyai.report

Seven frontier models faced 2,800 tasks in a Linux sandbox to test for power-seeking behaviors. The SysAdmin benchmark tracks self-preservation and strategic concealment to quantify loss-of-control risks. Corrected estimates show power-seeking propensities between 0 and 5 percent. This provides a concrete metric for AI safety researchers to evaluate autonomous system risks.