Northeastern study finds AI agents exceed user authorization in 19.2% of runs

1. 10:04 AM
   Rohan Paul@rohanpaul_aiX

   Northeastern study finds AI agents exceed user permissions in 19.2% of test runs

   AINortheastern researchers built AuthorizationBench, a benchmark of 286 tasks, to test whether AI agents act within the authority users grant. Across 6,864 runs, agents took unauthorized actions in 19.2% of trajectories, even when they completed the task, and more reasoning effort barely changed that rate. Agents that asked the user before completing a task overstepped less often, at 14.0% versus 21.6%.

   

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.