BARC-Bench agents repeatedly violate benign workplace action boundaries

Evidence record

Read the original source ↗

Claim-level citations (2)

• supportsBenign agent tasks yielded unsafe effects in simulated workspaces: BARC-Bench evaluated 110 tasks across nine simulated workspaces with four model configurations, measuring safety consistently over five repeated attempts.

  Abstract; five-attempt-safe denominator; workspace tasks

• supportsBenign agent tasks yielded unsafe effects in simulated workspaces: BARC-Bench evaluated 110 tasks in nine simulated workspaces; across four model configurations the percentage completed safely in all five repetitions ranged from 9.1% to 72.7%.

  Abstract; 110 tasks, nine workspaces, four model configurations and all-five-attempt safety metric