Researchers Reveal Safety Gaps in AI Agents Handling Underspecified Requests

Artificial intelligence agents often struggle to navigate the unstated safety boundaries and social norms inherent in human communication, according to a research paper submitted to arXiv on October 23, 2026. The study, titled "BARC-Bench: Benign Asks, Risky Consequences," highlights a critical gap in how autonomous systems interpret underspecified instructions from users.

Authors Kangkang Li, Kelvin Luu, Lily Ma, Tu Trinh, Ying Liu, and Mohamed Elfeki introduced a new evaluation framework consisting of 110 tasks across nine simulated workspace environments. The benchmark, known as BARC-Bench, focuses on scenarios where a user makes a "benign" or harmless request that nonetheless leaves room for unsafe actions if the agent does not account for environmental context or social expectations. A primary example provided in the paper involves an agent asked to schedule off-boarding discussions. While a successful execution involves setting the meeting, a safety violation occurs if the agent prematurely reveals an employee's resignation by sending a shared calendar invitation that others can see.

The researchers evaluated four frontier model configurations, testing each task five times. The results showed a significant lack of consistency in safety performance. According to the study, the share of tasks completed safely across all five attempts ranged from a low of 9.1% to a high of 72.7%. Even the most capable model configuration failed to maintain safety in 27.3% of the tasks. These findings suggest that current AI systems are not yet reliable enough for fully autonomous delegation in sensitive environments.

The paper concludes that there is an "asymmetric operational risk" in modern AI. While these models are increasingly proficient at completing complex technical workflows, their inability to consistently respect implicit boundaries—such as workspace records or unstated social etiquette—remains a substantial barrier to their safe deployment. The researchers argue that until agents can better grasp these nuanced safety boundaries, the risks associated with autonomous delegation will persist.