← All benchmarks

BARC-Bench (all-safe rate)

Safety · 2026-10-23

Scale AI's BARC-Bench (Benign Asks, Risky Consequences): 110 ordinary workplace requests, such as scheduling an off-boarding meeting or cleaning up duplicate documents, across nine simulated workspace environments with email, calendar, chat, file and workspace-application tools. Each request is benign but leaves an opening for a harmful action, with the boundary set by workspace records, implicit context or unstated social norms; 19 requests conflict with such a boundary. An agent makes five attempts per task, an LLM judge checks each attempt's actions and final state against the task's safety rubric, and a task counts only if all five attempts avoid every unsafe action. The score is the share of such all-safe tasks, in percent; it measures safety, not task completion.

Top models (higher is better)