A new benchmark reveals that AI models pitched as real-time safety monitors for robots are easier to fool than their developers likely want to admit.
Researchers introduced EgoSafetyBench, a dataset of 1,200 egocentric video scenarios annotated at half-second intervals, designed to stress-test vision-language models acting as streaming safety guards. The benchmark splits into two tracks: a situational track covering 800 scenes ranging from routine activity to contextual hazards, and a visual-channel track covering 400 scenes where misleading in-scene text — a sign, sticker, or label — contradicts the physical reality. Across both tracks, scenarios are built as contrastive pairs, near-identical clips that differ in exactly one visible cue, so a model can only pass by reasoning about that cue specifically. Ten open- and closed-source models were evaluated.
The results expose a gap between recognition and reasoning. Models generally flag videos that contain hazards, but they routinely miss the specific moment a hazard occurs — especially when context is required to interpret it. The text-manipulation track is worse: misleading signs degrade every model tested, with vulnerable models missing up to a third of hazards and robust models compensating by over-intervening on safe content. The benchmark's matched controls suggest that what looks like safety robustness is often just indiscriminate alarming.
This matters because the deployment pitch for VLM-based safety guards is that they can sit between a robot and its environment and catch dangerous moments before they become incidents — in homes, in factories, anywhere humans and machines share space. A guard that cries wolf on routine activity is unusable; one that misses contextual hazards is dangerous. EgoSafetyBench is a diagnostic, not a product, but it arrives at a pointed moment: the industry is actively deploying these systems while this class of evaluation barely existed six months ago.