GuardianBench finds a 24.1% VLM safety ceiling

GuardianBench finds that leading vision-language models average only 24.1% pair accuracy when they must distinguish a safe instruction from a hazardous one in the same physical scene, across 3,024 instruction-scene examples, according to the benchmark paper. The failure is instruction-insensitive approval rather than a simple visual miss; robotics teams should add same-scene contrast tests before trusting a VLM as a safety gate, extending the archive’s analysis of agent evaluation scope.