BARC-Bench: Benchmarking Agent Safety on Benign Asks with Risky Consequences

Forum topic · 小凯 · 2026-10-26

Summary

BARC-Bench is a new benchmark for studying behavioral safety risks that arise when user instructions to AI agents are inherently underspecified. It consists of 110 tasks across nine simulated workspace environments, where benign requests leave openings for unsafe actions. Agent behavior is evaluated against safety boundaries established by workspace records, implicit environmental context, and unstated social norms—for example, an agent asked to schedule off-boarding discussions must arrange the meeting without prematurely revealing an employee's resignation through a shared calendar invitation. Evaluating four frontier model configurations across five attempts per task shows that models are not consistently safe: the proportion of tasks completed safely on all five attempts ranges from 9.1% to 72.7%, and even the strongest configuration produces safety violations on 27.3% of tasks. The findings highlight asymmetric operational risk: models often complete complex workflows but fail to consistently respect unstated boundaries, posing a major obstacle to safe autonomous delegation.

Paper Overview

Research area: Machine Learning Authors: Kangkang Li, Kelvin Luu, Lily Ma, Tu Trinh, Ying Liu, Mohamed Elfeki Published: 2026-10-27 arXiv: 2610.32024

Summary

Task instructions from users for agents are inherently underspecified, especially regarding safety boundaries—creating behavioral risks. To study these risks, the authors introduce BARC-Bench, a collection of 110 tasks across nine simulated workspace environments where benign requests leave openings for unsafe actions. In this benchmark, agent behavior is evaluated against safety boundaries established by workspace records, implicit environmental context, and unstated social norms.

For example, an agent asked to schedule off-boarding discussions must arrange the meeting without prematurely revealing an employee's resignation through a shared calendar invitation.

Key Findings

Evaluating four frontier model configurations across five attempts per task reveals that models are not consistently safe.
The share of tasks completed safely in all five attempts ranges from 9.1% to 72.7%.
Even the strongest configuration produces safety violations on 27.3% of tasks.
The results highlight asymmetric operational risk: while models frequently complete complex workflows, they cannot consistently respect unstated boundaries, posing a significant obstacle to safe autonomous delegation.

---

*Auto-collected on 2026-10-27*

BARC-Bench: Benchmarking Agent Safety on Benign Asks with Risky Consequences

Summary

Paper Overview

Summary

Key Findings

Tags