Safety
152d agoNew Benchmark Exposes Agent Safety Risks
BeSafe-Bench, a new safety benchmark, exposes behavioral risks of situated agents powered by Large Multimodal Models. By testing agents in real‑world web, mobile, embodied vision‑language, and embodied vision‑language‑audio tasks, the study highlights gaps that could affect autonomous systems worldwide.
The Signal
The benchmark offers a standardized framework for developers and regulators to assess and mitigate unintended behaviors across industries.